Thread: Including Snapshot Info with Indexes

Hi,<br />    Currently The index implementation in Postgresql does not store the  Snapshot information in the Index. If
weadd the snapshot information into the indexing structure, we will have the following advantages.<br /><br /> a) There
canbe index only scans like Oracle<br />b) Unique indexes will become less costly, as older index tuples can be found
out.<br/>c) Even the index scans will get faster, since some of the index tuples won't translate into HeapScans. <br
/>d)Deletes and Updates will become slightly costly, as they have to update these indexes.<br /><br />I propose we add
aDDL like "Create Index .. With Snapshot", which will have different relkind.<br /><br />The design in my mind is to
addthe Snapshot info together with the values as first variables. We need to change the Index_getattr slightly to have
therelkind as input parameter, based on which we can have an offset. <br /><br />Please reply back with your
comments.<br/><br />Thanks,<br />Gokul.<br /> 

Gokulakannan Somasundaram wrote:
>     Currently The index implementation in Postgresql does not store the
> Snapshot information in the Index. If we add the snapshot information into
> the indexing structure, we will have the following advantages.

This idea has been discussed to death many times before. Please search
the archives.

> a) There can be index only scans like Oracle

IMO, the most promising approach to achieving index-only-scans at the
moment is the Dead Space Map, as discussed in the 8.3 dev cycle.

> b) Unique indexes will become less costly, as older index tuples can be
> found out.

Doesn't seem like a big benefit, considering that in most cases there
won't be any tuples in the index with a duplicate key. A common
exception to that is (non-HOT) updating a row. But in that case, the
page containing the old tuple is already in cache, so the lookup of the
visibility from the heap is cheap.

> c) Even the index scans will get faster, since some of the index tuples
> won't translate into HeapScans.

That's the same as doing an index-only-scan, right?

> d) Deletes and Updates will become slightly costly, as they have to update
> these indexes.

I think you're grossly underestimating the cost of that. For example, on
a table with 3 indexes. a delete currently requires one index lookup +
one heap lookup. With visibility in the indexes, that would require 3
index lookups + one heap lookup. That's 4 vs. 2 page accesses, not
taking into account the non-leaf b-tree pages. The real impact will
depend on what's in cache, but the cost can be very high.

Also, the full visibility information would need 12 bytes of space per
tuple. An index tuple on an int4 key currently takes 12 bytes, so that
would double the index size. Storage size has a big impact on
performance. More bytes means more I/O, less data fits in cache, and
more WAL traffic.

There's non-trivial implementation issues involved as well. You'd need a
way to reliably find all the index pointers for a given heap tuple
(search the archives for "retail vacuum" for the issues involved in
that. Broken user defined functions are a problem for example). And
you'd need to keep them all locked at the same time to modify them all
atomically, which is prone to deadlocks.

--  Heikki Linnakangas EnterpriseDB   http://www.enterprisedb.com

On Mon, 2007-10-08 at 09:40 +0100, Heikki Linnakangas wrote:
> This idea has been discussed to death many times before. Please search
> the archives.

Somewhat related to the "visibility in index" thing: would it be
possible to have a kind of index-table ? We do have here some tables
which have 2-4 fields, and the combination of them forms the primary key
of the table. There are usually no other indexes on the table, and the
net result is that the PK index duplicates the heap. So it would be cool
if the index would be THE heap somehow...

The most prominent example of this in our DBs is this table:

db> \d table_a Table "public.table_a" Column   |  Type  | Modifiers 
-----------+--------+-----------  id_1    | bigint | not null  id_2    | bigint | not null
Indexes:   "pk_table_a" PRIMARY KEY, btree (id_1, id_2)

db> select reltuples::bigint, relpages from pg_class where
relname='table_a';reltuples | relpages 
-----------+----------445286464 |   710090
(1 row)

db> select reltuples::bigint, relpages from pg_class where
relname='pk_table_a';reltuples | relpages 
-----------+----------445291072 |   599848
(1 row)

This postgres install is compiled with 32K page size (for the ones who
wonder about the page counts). In any case, it is clear that the index
basically duplicates the heap...

Cheers,
Csaba.

Csaba Nagy wrote:
> On Mon, 2007-10-08 at 09:40 +0100, Heikki Linnakangas wrote:
>> This idea has been discussed to death many times before. Please search
>> the archives.
> 
> Somewhat related to the "visibility in index" thing: would it be
> possible to have a kind of index-table ? We do have here some tables
> which have 2-4 fields, and the combination of them forms the primary key
> of the table. There are usually no other indexes on the table, and the
> net result is that the PK index duplicates the heap. So it would be cool
> if the index would be THE heap somehow...

The clustered index patch I worked on for 8.3, but didn't make it in,
would have helped that use case a lot.

A column-store kind of mechanism would be nice. Some columns could be
stored in index-like structures, while other would be in the heap. I
haven't seen any practical proposal on how to do it though.

--  Heikki Linnakangas EnterpriseDB   http://www.enterprisedb.com

Gokulakannan Somasundaram wrote:
> On 10/8/07, Heikki Linnakangas <heikki@enterprisedb.com> wrote:
>> IMO, the most promising approach to achieving index-only-scans at the
>> moment is the Dead Space Map, as discussed in the 8.3 dev cycle.
> 
> Index only scans means that  in order to get certain results, we may not
> goto the table at all. For example, if you have an index on columns a and b,
> and if there is a query like "select b from table where a between a1 and
> a2", then the explain plan need not goto the table. I can't understand how
> dead space map will provide such a functionality. 

Please read the discussions in the archives. The dead space map holds
visibility information in a condensed form. For index-only-scans, we
need to know if all tuples on page are are visible to us. If the dead
space map is designed with index-only-scans in mind, we can store a bit
there indicating "all tuples on this page are visible to everyone".
Pages that have that bit set don't need to be visited to check visibility.

What information exactly is going to be stored in the dead space map is
still debated. For vacuuming, we need to know which pages contain dead
tuples that are worth vacuuming, which isn't the same thing as "all
tuples are visible to everyone", but it's quite close.

Heap pages that do have dead or recently modified rows do need to be
visited, so the executor needs to always be prepared to check visibility
from the heap. However, on a table that's not very frequently updated,
most bits will be set and the effect will be the same as genuine
index-only-scans. On a table that is frequently updated, you would
suffer a big hit in update performance with the "duplicate visibility
information in all indexes" scheme as well, as the updates would need to
update the indexes as well, so the performance characteristics are
roughly the same.

> That's true. But i am not asking to replace the current index
> implementation, but to provide an extra option while indexing. Say if a
> particular database setup doesn't do much deletes and updates(imagine tables
> with partitioning, where the partitions/tables are dropped instead of
> deletes. They can have an option to "create index .. with snapshot"

A nice property of utilizing the dead space map for index-only-scans is
that it doesn't require any action from the DBA. It will "just work". It
also adapts well to tables that have parts that are frequently updated,
and other parts that are not. The frequently updated parts will behave
like what we have now, index-only-scans are not possible because, but
deletes/updates are cheap. But the less frequently updated parts will
eventually have the bits set in the map, and we can do index-only-scans
for those parts.

> Imagine the Index Vacuum also will do lesser Random I/Os

Index vacuum doesn't do random I/Os as it is.

> Also, the full visibility information would need 12 bytes of space per
>> tuple. An index tuple on an int4 key currently takes 12 bytes, so that
>> would double the index size. Storage size has a big impact on
>> performance. More bytes means more I/O, less data fits in cache, and
>> more WAL traffic.
> 
> I am thinking of certain optimizations here.  we  have a bit unused in
> indextuple structure.  If a particular tuple is not deleted, then we can
> signify that using that bit and save 6 bytes of saving the xmax and cmax. We
> are trading of this space efficiency in place of Random I/Os, which is not a
> bad trade-off , i suppose. Again this is going to optional for the user. If
> users have an option to create Bitmap index/ Binary index, why can't they
> have this option as well?

Because we have to maintain it? :)

Speaking of bitmap indexes, that would be nice to have. It looks like
Gavin dropped the ball on the patch, so if you want to continue that
work, that would be great.

> There's non-trivial implementation issues involved as well. You'd need a
>> way to reliably find all the index pointers for a given heap tuple
>> (search the archives for "retail vacuum" for the issues involved in
>> that. Broken user defined functions are a problem for example). And
>> you'd need to keep them all locked at the same time to modify them all
>> atomically, which is prone to deadlocks.
> 
> I think Vacuum need not goto the table, as the visibility information is
> present in the index itself. 

Vacuum isn't the problem here. The problem is: given heap tuple X, how
do you find the the corresponding index tuples? The obvious solution is
to calculate the index keys from the heap tuple, and use them to look
up. But what if you have an expression index on a user-defined function,
and that user-defined function is broken so that it returns a different
value than when the index tuple was inserted? You won't find the index
tuples in that case, so you won't be able to update the visibility
information. Granted, you've got a broken user-defined-function in that
case, and you're going to get bogus query results anyway. But not
finding the index tuple when needed would lead to more serious
corruption. This is the same problem many proposed retail vacuum schemes
have stumbled into, which is why I suggested searching for that.

--  Heikki Linnakangas EnterpriseDB   http://www.enterprisedb.com

Ühel kenal päeval, E, 2007-10-08 kell 11:41, kirjutas Heikki
Linnakangas: 
> The dead space map holds
> visibility information in a condensed form. For index-only-scans, we
> need to know if all tuples on page are are visible to us. If the dead
> space map is designed with index-only-scans in mind, we can store a bit
> there indicating "all tuples on this page are visible to everyone".
> Pages that have that bit set don't need to be visited to check visibility.
> 
> What information exactly is going to be stored in the dead space map is
> still debated. For vacuuming, we need to know which pages contain dead
> tuples that are worth vacuuming, which isn't the same thing as "all
> tuples are visible to everyone", but it's quite close.

I would prefer a separate MVC visibility heap (aka. extended "dead space
map") which would duplicate whole visibility info from heap pages, just
in compressed form. After a few releases with duplicated visibility
info, we could remove it from the data heap.

If the whole visibility info (cmin, cmax, tmin, tmax, flags, (+ size for
DSM uses)) for tuples, were in a separate heap, it would allow for a lot
of on-the-fly compression. for example we could:

* get rid of both tmin and tmax for all completed transactions
* reduce any deleted tuple to just flags
* reduce any tuple produced by aborted transaction to just flags
* reduce any tuple visible to all backends to just flags
* RRL compress (runs of) pages produced by same transaction 
* RRL compress (runs of) pages with all tuples visible
* RRL compress (runs of) pages with all tuples deleted

depending on distribution of Inserts/Updates/Deletes it will make
visibility info a little or a lot smaller than it is currently, greatly
enchancing chances that it stays in cache (even for OLAP loads, because
data for these are usually produced by bulk inserts and thus their
visibility info is  highly compressable)

It also makes VACUUM more efficient, as it's initial scan (find
vacuumable tuples) will need to do lot less IO.

And it allows for more intelligent choices for new tuple placement ,
especially if we want to preserve clustering.

And of course it gives you kind of index-only scans (mostly read index + check in vis.heap)

-------------
Hannu

Ühel kenal päeval, E, 2007-10-08 kell 12:12, kirjutas Gokulakannan
Somasundaram:
> Hi,
>     Currently The index implementation in Postgresql does not store
> the  Snapshot information in the Index.
..
> Please reply back with your comments.

I think you got enoght "search for previous discussion" answers on this
aone already ;) 

So I propose a few another ideas to investigate

1. get rid of indexes for TOAST tables

instead of TOAST tuple id store CTID of first TOAST block directly, and
use something like skip lists inside the TOAST block headers to access
next TOAST tuples.

2. store visibility info in TOAST tables

if you store visibility info in TOAST tuples, it becomes possible to
update just the small part of the tuple in the main heap and keep the
bulk of toasted data in place.


both of these ideas are much more complicated to implement than it
appears from my simple description, but should have big benefits for a
sizable number of scenarios which use TOAST.

-------------
Hannu

On Mon, 2007-10-08 at 14:58 +0300, Hannu Krosing wrote:

> 1. get rid of indexes for TOAST tables
> 
> instead of TOAST tuple id store CTID of first TOAST block directly, and
> use something like skip lists inside the TOAST block headers to access
> next TOAST tuples.

It should be possible to optimise TOAST for when there is just a single
chunk that is toasted. Since we often (and by default) compress data
stored in TOAST tables this would be a frequently used optimisation. 

Instead of storing the TOAST OID, which is then looked-up in the index
to find the TOAST tid, we can just store the tid directly in the toast
pointer on the main heap. That way we'll get faster read and write
access for a good proportion of rows by avoiding the toast index and
going straight to the toast heap.

We'd need a different kind of toast pointer which would be 2 bytes
longer than the normal pointer. I think we have a spare flag bit to
indicate this.

That's just a rough sketch, I've not checked the details on this.

--  Simon Riggs 2ndQuadrant  http://www.2ndQuadrant.com

Hi Heikki,
             I am always slightly late in understanding things. Let me try to understand the use of DSM. It is a bitmap index on whether all the tuples in a particular block is visible to all the backends, whether a particular block contains tuples which are invisible to everyone. But i think this will get subjected to the same limitations of Bitmap index. Even Oracle suggests the use of Bitmap index for only data warehousing tables, where the Bitmap indexes will be dropped and recreated after every bulk load. This is not a viable alternative for OLTP  transactions. But i think i am late in the game as i haven't participated in those discussions
           One Bitmap index block usually maps to lot of blocks in the heap. So locking of one page to update the DSM for update/delete/insert would hit the concurrency. But again all these are my observation w.r.t oracle bitmap indexes. May be i am missing something in DSM.
         I couldn't get that piece of discussion in the archive, which discusses the design of Retail Vacuum. So please advise me again here. 
          Let's take up Retail Vacuuming again. The User defined function which would return different values at different time can be classified as non-deterministic  functions. We can say that this index cannot be created on a non-deterministic function. This is the way it is implemented in Oracle. What they have done is they have classified certain built-in operators and functions as deterministic. Similarly they have classified a few as non-deterministic operators and functions. Can we  follow a similar approach?
         

Thanks,
Gokul.
        
 

Gokulakannan Somasundaram wrote:
> Hi Heikki, I am always slightly late in understanding things. Let me
> try to understand the use of DSM. It is a bitmap index on whether all
> the tuples in a particular block is visible to all the backends,
> whether a particular block contains tuples which are invisible to
> everyone. But i think this will get subjected to the same limitations
> of Bitmap index. Even Oracle suggests the use of Bitmap index for
> only data warehousing tables, where the Bitmap indexes will be
> dropped and recreated after every bulk load. This is not a viable
> alternative for OLTP  transactions. But i think i am late in the game
>  as i haven't participated in those discussions
While the DSM might be similar in spirit to a bitmap index, the actual
implementation has a lot more freedome I'd say, since you can tailor it
exactly to the need of tracking some summarized visibility info. So not
all shortcomings of bitmap indices must necessarily apply to the DSM
also. But of course thats mostly handwavering...

> One Bitmap index block usually maps to lot of blocks in the heap. So
> locking of one page to update the DSM for update/delete/insert would
> hit the concurrency. But again all these are my observation w.r.t
> oracle bitmap indexes. May be i am missing something in DSM.
A simple DSM would probably contain a bit per page that says "all xmin <
GlobalXmin, and all xmax unset or aborted". That bit would only get SET
during VACUUM, and only unset during INSERT/UPDATE/DELETE. If setting it
is protected by a VACUUM-grade lock on the page, we might get away with
no locking during the unset, making the locking overhead pretty small.


> I couldn't get that piece of discussion in the archive, which 
> discusses the design of Retail Vacuum. So please advise me again
> here. Let's take up Retail Vacuuming again. The User defined function
>  which would return different values at different time can be
> classified as non-deterministic  functions. We can say that this
> index cannot be created on a non-deterministic function. This is the
> way it is implemented in Oracle. What they have done is they have
> classified certain built-in operators and functions as deterministic.
> Similarly they have classified a few as non-deterministic operators
> and functions. Can we  follow a similar approach?
Postgres already distinguishes VOLATILE,STABLE and IMMUTABLE functions.
It doesn't, however, risk physical data corruption, even if you get that
classification wrong. The worst that happens AFAIK are wrong query
results - but fixing your function, followed by a REINDEX always
corrects the problme. If you start poking holes into that safety net,
there'll be a lot of pushback I believe - and IMHO rightly so, because
people do, and always will, get such classifications wrong.

greetings, Florian Pflug

Gokulakannan Somasundaram wrote:
>              I am always slightly late in understanding things. Let me try
> to understand the use of DSM. It is a bitmap index on whether all the tuples
> in a particular block is visible to all the backends, whether a particular
> block contains tuples which are invisible to everyone. But i think this will
> get subjected to the same limitations of Bitmap index. Even Oracle suggests
> the use of Bitmap index for only data warehousing tables, where the Bitmap
> indexes will be dropped and recreated after every bulk load. This is not a
> viable alternative for OLTP  transactions. 

Well, it's not quite the same as a bitmap index, though both use a
bitmap. You didn't quite get into details on what the limitations are
and why it wouldn't be suitable for OLTP, but I don't see any
significant problems.

> But i think i am late in the game
> as i haven't participated in those discussions

Better late than never :).

>            One Bitmap index block usually maps to lot of blocks in the heap.
> So locking of one page to update the DSM for update/delete/insert would hit
> the concurrency. But again all these are my observation w.r.t oracle bitmap
> indexes. May be i am missing something in DSM.

Yeah, the DSM page could become a contention bottleneck. My current
thinking is that we'd have a flag in the heap page header, that would be
set together with the bit in the DSM. When the flag in the page header
is set, you don't need to lock and update the DSM because you know the
bit is already set. Vacuum would have to clear both the DSM bit and the
flag.

>           Let's take up Retail Vacuuming again. The User defined function
> which would return different values at different time can be classified as
> non-deterministic  functions. We can say that this index cannot be created
> on a non-deterministic function. This is the way it is implemented in
> Oracle. What they have done is they have classified certain built-in
> operators and functions as deterministic. Similarly they have classified a
> few as non-deterministic operators and functions. Can we  follow a similar
> approach?

We already do. A function must be marked as IMMUTABLE in order to use it
in an index expression. But we can't enforce that the user defined
function really behaves like an immutable function should. If someone
creates a user-defined function in C that calls the C random() function,
we can't stop it.

As I said earlier, using an index like that will of course lead to bogus
results. But it won't currently cause any server crashes or more serious
corruption.

--  Heikki Linnakangas EnterpriseDB   http://www.enterprisedb.com

"Gokulakannan Somasundaram" <gokul007@gmail.com> writes:

> On 10/9/07, Gokulakannan Somasundaram <gokul007@gmail.com> wrote:
>
>> A function is said to be deterministic, if it returns the same value,
>> irrespective of how many times, it is invoked. I think this definition
>> clearly puts the random function under the non-deterministic category. If we
>> have such a classification, do you think we can resolve this issue?
>
> If we frame a set of guidelines/test procedure, do you think it might solve
> the issue? Even, if we don't allow this type of indexing to anything other
> than built-in deterministic functions, i feel it would serve most of the
> indexing requirements.

We already do this. c.f. IMMUTABLE at

http://www.postgresql.org/docs/8.2/interactive/xfunc-volatility.html

and

http://www.postgresql.org/docs/8.2/interactive/sql-createindex.html


--  Gregory Stark EnterpriseDB          http://www.enterprisedb.com

[snip]
> In the case of User-Defined functions, the user should be defining it
> as Deterministic. 

The user CAN already define his functions as
"Deterministic=IMMUTABLE"... the problem is that many of us will define
functions as immutable, when in fact they are not. And do that by
mistake... and there's nothing postgres can do about that.

> Can we frame a set of guidelines, or may be some test procedure, which
> can declare a certain function as deterministic? 

You mean postgres should check your function if it is really immutable ?
I can't imagine any way to do it correctly in reasonable time :-)
Imagine a function of 10 parameters which returns the sum of the
parameters all the time except for parameters all 1 it will randomly
return a value _once in a thousand executions_... please find a generic
algorithm which spots this function as not immutable in reasonable
execution time ;-)
So this example is a bit extreme, but don't underestimate the user ;-)

> I am just saying from the top of my mind. Even otherwise, if we can
> even restrict this indexing to only Built-in deterministic functions.,
> don't you think it would help the cause of a majority? I have just
> made the proposal to create the index with snapshot a optional one. 

Restrictions like this are always confusing for the end user (i.e. why
can I use built-ins here and not my own ?). I leave to the actual coders
to say anything about code maintenance concerns...

Cheers,
Csaba.

Csaba Nagy wrote:
>> Can we frame a set of guidelines, or may be some test procedure, which
>> can declare a certain function as deterministic? 
> 
> You mean postgres should check your function if it is really immutable ?
> I can't imagine any way to do it correctly in reasonable time :-)
> Imagine a function of 10 parameters which returns the sum of the
> parameters all the time except for parameters all 1 it will randomly
> return a value _once in a thousand executions_... please find a generic
> algorithm which spots this function as not immutable in reasonable
> execution time ;-)
> So this example is a bit extreme, but don't underestimate the user ;-)

I think you're overly pessimistic here ;-) This classification can be done quite 
efficiently as long as your language is "static enough". The trick is not to 
execute the function, but to scan the code to find all other functions and SQL 
statements a given function may possibly call. If your function calls no SQL 
statements, and only other functions already marked IMMUTABLE, then it must be 
IMMUTABLE itself.

It does seem that only pl/pgsql is "static enough" for this to work, though,
making this idea rather unappealing.

>> I am just saying from the top of my mind. Even otherwise, if we can
>> even restrict this indexing to only Built-in deterministic functions.,
>> don't you think it would help the cause of a majority? I have just
>> made the proposal to create the index with snapshot a optional one. 
> 
> Restrictions like this are always confusing for the end user (i.e. why
> can I use built-ins here and not my own ?). I leave to the actual coders
> to say anything about code maintenance concerns...
Yes, and some built-ins have gotten that classification wrong too in the past 
IIRC. Which probably is a good reason not to trust our users to get it right ;-)

greetings, Florian Pflug

> I think you're overly pessimistic here ;-) This classification can be done quite 
> efficiently as long as your language is "static enough". The trick is not to 
> execute the function, but to scan the code to find all other functions and SQL 
> statements a given function may possibly call. If your function calls no SQL 
> statements, and only other functions already marked IMMUTABLE, then it must be 
> IMMUTABLE itself.

OK, I have a "black-box" mindset right now due to the problem I'm
currently working on, so I didn't even think about checking the source
code of the function (which is the right thing to do if you have the
source code)... in which case you're right, I was overly pessimistic :-)

Cheers,
Csaba.

Csaba Nagy wrote:
> You mean postgres should check your function if it is really immutable ?
> I can't imagine any way to do it correctly in reasonable time :-)
>   
>   

I would say that in the general case it's analogous to the halting 
problem, not solvable at all let alone in any reasonable time.

cheers

andrew

Florian G. Pflug wrote:
>
> I think you're overly pessimistic here ;-) This classification can be 
> done quite efficiently as long as your language is "static enough". 
> The trick is not to execute the function, but to scan the code to find 
> all other functions and SQL statements a given function may possibly 
> call. If your function calls no SQL statements, and only other 
> functions already marked IMMUTABLE, then it must be IMMUTABLE itself.
>
> It does seem that only pl/pgsql is "static enough" for this to work, 
> though,
> making this idea rather unappealing.
>
>

How would you propose to analyse C functions, for which you might not 
have the C code?

cheers

andrew

On Tue, 2007-10-09 at 11:22 -0400, Andrew Dunstan wrote:
> 
> Csaba Nagy wrote:
> > You mean postgres should check your function if it is really immutable ?
> > I can't imagine any way to do it correctly in reasonable time :-)

> I would say that in the general case it's analogous to the halting 
> problem, not solvable at all let alone in any reasonable time.

In the light of Florian's mail, I would say that in the context of a
language which can check each of it's constructs if it is immutable or
not, a procedure using only immutable constructs should be itself
immutable... the halting problem is avoided in that you don't really
need to know if/how the procedure works, you only need to know that it
will always work the same ;-) The problem is that in the general case
the languages don't have available checks for this kind of thing, so
either you restrict the immutability check to simple languages ("static
enough" as Florian would say) or you must allow the user to decide if
the function is immutable or not. In the general case I assume the users
will want the power to decide (and potentially be wrong), and will
expect that if they do mistake, the result won't be catastrophic. I
guess this is the same conclusion as in previous threads about the
subject...

Cheers,
Csaba.

Andrew Dunstan wrote:
> Florian G. Pflug wrote:
>> 
>> I think you're overly pessimistic here ;-) This classification can be done
>> quite efficiently as long as your language is "static enough". The trick is
>> not to execute the function, but to scan the code to find all other
>> functions and SQL statements a given function may possibly call. If your
>> function calls no SQL statements, and only other functions already marked
>> IMMUTABLE, then it must be IMMUTABLE itself.
>> 
>> It does seem that only pl/pgsql is "static enough" for this to work, 
>> though, making this idea rather unappealing.
>> 
> 
> How would you propose to analyse C functions, for which you might not have
> the C code?
Scanning the binary, together with symbol annotations for immutability of course
;-))

No, seriously. I do *not* advocate that we actually autoclassify functions, for
a lot of reasons. I just wanted to refute the statement that doing so is
generally impossible - it's not. It's trivial for some languages (In haskhell
for example all functions that don't use monads are immutable, and their
signature tell if they do use monads or or), realistic for others (pl/pgsql,
where we do have the sourcecode), and utterly impossible for others
(pl/{ruby,python,perl,...}, pl/c, ...).

Besides - AFAICS *anything* that makes VACUUM depend on IMMUTABLE to be correct
would instantly break tsearch, no? At least as long as we allow changing
stopwords and the like of dictionaries used by an index - which we'd better
allow, unless we want the DBAs to come with pitchforks after us...

regards, Florian Pflug, who shudders when imagining DBAs with pitchforks...

Gokulakannan Somasundaram wrote:
>  As explained, if we are going to include the snapshot with indexes, Vacuum
> will be done on the index independent of the table, so Vacuum will not
> depend on immutability. We need to goto the index from the table, when we
> want to update the snapshot info. The problem on hand is that some of the
> userdefined functions are mutable, whereas the user might mark it immutable.
> 
> So my idea is to have a mapping index, with tupleid as the first column and
> the function's values as subsequent columns. I have a somewhat detailed
> design in mind. So there will be a over head of extra 3 I/Os for
> update/delete on indices based on User-defined functions. But this setup
> will speed-up lot of queries where the tables are partitioned and there will
> be more inserts and selects and dropping partitions at periodic intervals.
> Updates become costly by 3 I/Os per Index with snapshot. So if someone has
> more selects than updates+deletes then this index might come handy (ofcourse
> not with user-defined functional indices).

I think you need to explain why that is better than using the Dead Space
Map. We're going to want the DSM anyway, to speed up VACUUMs; enabling
index-only-scans just came as an afterthought. While DSM designed just
for speeding up vacuums might look slightly different than one used for
index-only scans, the infrastructure is roughly the same.

What you're proposing sounds a lot more complex, less space-efficient,
and slower to update. It requires extra action from the DBA, and it
covers exactly the same use case (more selects than updates+deletes, to
use your words). It would require changes to all index access methods,
while the DSM would automatically work with all of them. In particular,
including visibility information in a bitmap index, should we have
bitmap indexes in the future, is impossible, while the DSM approach
would just work.

--  Heikki Linnakangas EnterpriseDB   http://www.enterprisedb.com

Hi All,
         So i think we are clear now, that it is possible to have an index with snapshot info.  Let me try to enumerate the uses of having the Index with snapshot info, in comparison to the Dead Space Map.

a) Dead Space, if it is successfull in its implementation of what it claims, will have the means to point out that all the tuples of certain chunks are frozen for registered relations and registered chunks. There would be lot of blocks which won't fall under this category.
              i) For example, if the records are newly inserted, that block can't be marked as containing all frozen tuples.

Hi All,
         Last mail was sent by mistake without completion. I apologize for that. i am continuing on that.

So i think we are clear now, that it is possible to have an index with snapshot info.  Let me try to enumerate the uses of having the Index with snapshot info, in comparison to the Dead Space Map.

a) Dead Space, if it is successfull in its implementation of what it claims, will have the means to point out that all the tuples of certain chunks are frozen for registered relations and registered chunks. There would be lot of blocks which won't fall under this category.
              i) For example, if the records are newly inserted, that block can't be marked as containing all frozen tuples.
             ii) Imagine the case where there is a batch job / Huge select query running in a enterprise for more than 6hrs. All the blocks which have got inserted into the tables, during this period might not be able to get the advantage of DSM

Hi All,
         Last two mails were sent by mistake without completion. I couldn't curse my system any further
         I apologize for that.
        
         If we comeback to the topic of discussion

So i think we are clear now, that it is possible to have an index with snapshot info.  Let me try to enumerate the uses of having the Index with snapshot info, in comparison to the Dead Space Map.

a) Dead Space, if it is successfull in its implementation of what it claims, will have the means to point out that all the tuples of certain chunks are frozen for registered relations and registered chunks. There would be lot of blocks which won't fall under this category.
              i) For example, if the records are newly inserted, that block can't be marked as containing all frozen tuples.
             ii) Imagine the case where there is a batch job / Huge select query running in a enterprise for more than 6hrs. All the blocks which have got inserted into the tables, during this period might not be able to get the advantage of DSM
            iii) Imagine the case for which i am actually proposing the Index with Snapshot infos. Partitioned tables. Every time a new table gets created, it has to get registered into the Deadspace. This requires more maintenance on the DBA Side
            iv) I understand the DeadSpaceLock to be a Global lock(If one transaction is updating the dead space for any unregistered chunk, no one else can query the DeadSpace). If my statement is right, then partitioned tables might not be able to benefit from DSM. We have to remember for tables with daily partitions, this would prove to be a nightmare
           

Other than that there are lot of advantages, i foresee with including the indexes with snapshots
i) Vacumming of these indexes need not be done with SuperExclusive Locks. It is possible to design a strategy to vacuum these indexes with Exclusive locks on pages
ii) The above would mean that index can be in operation while the vacuum is happening
iii) As we have already stated, it provides a efficient clustering of related data.
iv) The Reverse Mapping Index, if present provides an efficient solution to the Retail Vacuum problem. So HOT can be improved further with no need to place the restriction of the updated tuple should be in the same page
iv) Updates on tables with primary keys(where primary key is not updated), will be able to resolve the unique constraint faster. This is a minor advantage.


The complexity of Reverse Mapping index will only be there for user-defined functional indexes.
These functions can be pruned further to find out the obvious immutable ones.


I expect your valuable feedback for this.

Thanks,
Gokul.

Gokulakannan Somasundaram wrote:
>          Last two mails were sent by mistake without completion. I couldn't
> curse my system any further

:-)

> a) Dead Space, if it is successfull in its implementation of what it claims,
> will have the means to point out that all the tuples of certain chunks are
> frozen for registered relations and registered chunks. There would be lot of
> blocks which won't fall under this category.
>               i) For example, if the records are newly inserted, that block
> can't be marked as containing all frozen tuples.

If records have just been inserted to a block, it is in cache. Therefore
hitting that block to check visibility isn't going to cost much. There
might be some middle-ground where a tuple has been inserted a while ago,
so that the block has already been evicted from cache, but the
transaction hasn't yet been committed, but that's a pretty narrow use case.

Note that we can flag a page in the DSM not only by VACUUM, but by any
backend as soon as all tuples are visible to everyone. You do have to
scan the tuple headers on the page to determine that, but that's not so
much overhead, and it could be offloaded to the bgwriter.

>              ii) Imagine the case where there is a batch job / Huge select
> query running in a enterprise for more than 6hrs. All the blocks which have
> got inserted into the tables, during this period might not be able to get
> the advantage of DSM

Yep, true. A long-running transaction like that is problematic anyway,
because we can't vacuum away any dead rows generated during that period.

>             iii) Imagine the case for which i am actually proposing the
> Index with Snapshot infos. Partitioned tables. Every time a new table gets
> created, it has to get registered into the Deadspace. This requires more
> maintenance on the DBA Side

Why do you think that the DBA needs to register tables to the DSM
manually? Surely that would happen automatically.

>             iv) I understand the DeadSpaceLock to be a Global lock(If one
> transaction is updating the dead space for any unregistered chunk, no one
> else can query the DeadSpace). If my statement is right, then partitioned
> tables might not be able to benefit from DSM. We have to remember for tables
> with daily partitions, this would prove to be a nightmare

The patch submitted for 8.3 did use a global lock, and a fixed size
shared memory area, but those were exactly the reasons the patch was
rejected. It will have to be reworked for 8.4.

> Other than that there are lot of advantages, i foresee with including the
> indexes with snapshots
> i) Vacumming of these indexes need not be done with SuperExclusive Locks. It
> is possible to design a strategy to vacuum these indexes with Exclusive
> locks on pages

I'm not convinced that's true. We only need super-exclusive locks on
index pages for interlocking index and heap accesses with non-MVCC
snapshots, IOW system tables. And since the lock is only held for a
short time and infrequently, it hasn't been a problem at all.

> ii) The above would mean that index can be in operation while the vacuum is
> happening

Huh? VACUUM hasn't locked out other access since version 7.2!

> iii) As we have already stated, it provides a efficient clustering of
> related data.

Sorry, I missed that part. What's that?

> iv) The Reverse Mapping Index, if present provides an efficient solution to
> the Retail Vacuum problem. So HOT can be improved further with no need to
> place the restriction of the updated tuple should be in the same page
> iv) Updates on tables with primary keys(where primary key is not updated),
> will be able to resolve the unique constraint faster. This is a minor
> advantage.
> 
> The complexity of Reverse Mapping index will only be there for user-defined
> functional indexes.

The *run-time* complexity of that will only be there for UDF indexes,
but the *code* complexity will always be there. Sorry, I think the
probability of a reverse mapping index being accepted to Postgres is
very close to zero.

--  Heikki Linnakangas EnterpriseDB   http://www.enterprisedb.com

Will $SUBJECT make it possible for count(*) to use index? This is a much
wanted feature.

--
Andreas Joseph Krogh <andreak@officenet.no>
Senior Software Developer / Manager
------------------------+---------------------------------------------+
OfficeNet AS            | The most difficult thing in the world is to |
Karenslyst Allé 11      | know how to do a thing and to watch         |
PO. Box 529 Skøyen      | somebody else doing it wrong, without       |
0214 Oslo               | comment.                                    |
NORWAY                  |                                             |
Tlf:    +47 24 15 38 90 |                                             |
Fax:    +47 24 15 38 91 |                                             |
Mobile: +47 909  56 963 |                                             |
------------------------+---------------------------------------------+

Andreas Joseph Krogh wrote:
> Will $SUBJECT make it possible for count(*) to use index? This is a much 
> wanted feature.

Yes, both the DSM approach and the approach proposed by Gokul.

--  Heikki Linnakangas EnterpriseDB   http://www.enterprisedb.com

On Friday 12 October 2007 11:49:17 Heikki Linnakangas wrote:
> Andreas Joseph Krogh wrote:
> > Will $SUBJECT make it possible for count(*) to use index? This is a much
> > wanted feature.
>
> Yes, both the DSM approach and the approach proposed by Gokul.

Good.

--
Andreas Joseph Krogh <andreak@officenet.no>
Senior Software Developer / Manager
------------------------+---------------------------------------------+
OfficeNet AS            | The most difficult thing in the world is to |
Karenslyst Allé 11      | know how to do a thing and to watch         |
PO. Box 529 Skøyen      | somebody else doing it wrong, without       |
0214 Oslo               | comment.                                    |
NORWAY                  |                                             |
Tlf:    +47 24 15 38 90 |                                             |
Fax:    +47 24 15 38 91 |                                             |
Mobile: +47 909  56 963 |                                             |
------------------------+---------------------------------------------+

Andreas Joseph Krogh <andreak@officenet.no> writes:
> Will $SUBJECT make it possible for count(*) to use index? This is a much
> wanted feature.

If you mean "count(*) will become instantaneous", no it won't.  It would
get faster, but probably not by more than a factor of 10 or so,
corresponding to the size ratio between index and table.  Remember that
the proposal is going to bloat indexes pretty badly: a simple index on
an int4 column will double in size, more or less.  So that ratio will
be much less favorable than it is today.

Personally I think the bloat problem will doom this entire approach.
The distributed costs of that will outweigh the speedups that can be
achieved for specific queries.  The OP is free to try to prove this
fear wrong, but no amount of discussion will constitute such a proof;
only a testable implementation.
        regards, tom lane

I agree with that. I will go ahead and do a test implementation and share the results with everyone.

Thanks,
Gokul.

Hi Simon/Hannu,
     
a) I accept with storing ctid in place of oid. That would definitely improve the performance. I guess it would also remove the restriction of the size on TOASTABLE ATTRIBUTES. I guess different chunks have to be linked together, without the index.

b) But i don't understand how storing the visibility info in TOAST table would help you. In that case you may need to update it for every delete/ update happening in the main table. Only thing, it would help is if someone wants to do a full table scan on TOAST table alone. May be Vacuum of TOAST tables can be done independently. But do you think it is worth the loss of performance in Update/Delete?

Thanks,
Gokul.

 

On Sat, 2007-10-13 at 11:12 +0530, Gokulakannan Somasundaram wrote:

> b) But i don't understand how storing the visibility info in TOAST
> table would help you. In that case you may need to update it for every
> delete/ update happening in the main table. Only thing, it would help
> is if someone wants to do a full table scan on TOAST table alone. May
> be Vacuum of TOAST tables can be done independently. But do you think
> it is worth the loss of performance in Update/Delete? 

No

--  Simon Riggs 2ndQuadrant  http://www.2ndQuadrant.com

"Gokulakannan Somasundaram" <gokul007@gmail.com> writes:

> I accept that the indexes will be bigger in size for this approach. You
> might need more disk-space  and you might need more memory to accomodate the
> same amount of information. But i think disk costs and memory costs have
> come down a lot, People can afford to buy more disk and memory. 

That's not how it works. We're not generally worried about people running out
of disk or memory resources. But no matter how cheap they get people will only
have what they have. We have to worry about running as fast as possible for a
*given* amount of RAM or disk.

Generally raising disk space usage results in a corresponding increase in run
time. So an index that takes twice as much space on disk will consume twice as
much time to consult as one that doesn't. You need to save enough time
elsewhere to make that up and then some to make it worthwhile.

I think we are pretty set on having the DSM for vacuuming purposes so you'll
also have to argue this approach will cover enough additional cases or be
better in some other way compared to using the DSM to be a win.

--  Gregory Stark EnterpriseDB          http://www.enterprisedb.com

Hi,
      I went through this article and it was good. Please have a look at it.

http://www.databasecolumn.com/2007/09/one-size-fits-all.html

This article was written by Michael Stonebraker, considered to be the founder of our database. He has mentioned that the DBMS designed in 1970s haven't changed according to the change that has happened in Hardware landscape. The Vertica database(Monet is a open source version with the same principle) makes use of the very same principle. Use more disk space, since they are less costly and optimize the data warehousing.

Even otherwise we are recommending Indexes with snapshot as an option. We are not replacing the current index scheme. So if someone feels that his database should run on lesser disk space, let them create the normal index. If he feels he can afford to have more redundant disk space, then he can create indexes with snapshots. We are reducing random I/Os at the cost of extra disk space. So definitely that's a good. But tech folks like us can better decide on something based on experiments, as Tom has pointed out. So let's see whether Indexes with snapshot is worth the trade-off in space.

Thanks,
Gokul.

On 10/13/07, Gokulakannan Somasundaram <gokul007@gmail.com> wrote:
> I accept that the indexes will be bigger in size for this approach. You
> might need more disk-space  and you might need more memory to accomodate the
> same amount of information. But i think disk costs and memory costs have
> come down a lot, People can afford to buy more disk and memory. But the only
> fact that remains true is that the disk, the last mechanical device is slow
> in addressing Random I/Os. So this proposal is aimed  at occupying more
> memory and disk space to reduce Random I/Os. Say, if we are accomodating 200
> tuples per index page in today's index(for a typical configuration), and as
> you said in the worst case (only one index field), we will be occupying 100
> tuples per index page. But we would take away probably 100 random I/Os (say
> with bitmap scan it reduces to 75). 1GB of memory is around $100 and 1GB of
> disk is around $1. But one hour of Performance tuner would cost around $200
> :)). So that's the trade-off for the enterprises, if they want to shift
> between the two indexes.

I disagree.  Even with the latest on-disk size enhancements, Postgres
still has a substantially larger on-disk footprint than pretty much
every other database.  Like it or not, additional I/O costs are not
something that should just be dismissed.

Disregarding fundamental database issues (like increased I/O) leads me
to believe that you don't have much experience tuning databases.  As I
have a bit of experience adding visibility to the indexes, I stand
behind DSM.  From an analytical standpoint, and given Postgres' MVCC
design, it would be hard to beat a properly designed DSM in this area.

-- 
Jonah H. Harris, Sr. Software Architect | phone: 732.331.1324
EnterpriseDB Corporation                | fax: 732.331.1301
499 Thornall Street, 2nd Floor          | jonah.harris@enterprisedb.com
Edison, NJ 08837                        | http://www.enterprisedb.com/

On 10/13/07, Gokulakannan Somasundaram <gokul007@gmail.com> wrote:
> Hi,
>       I went through this article and it was good. Please have a look at it.
>
> http://www.databasecolumn.com/2007/09/one-size-fits-all.html
>
> This article was written by Michael Stonebraker, considered to be the
> founder of our database. He has mentioned that the DBMS designed in 1970s
> haven't changed according to the change that has happened in Hardware
> landscape. The Vertica database(Monet is a open source version with the same
> principle) makes use of the very same principle. Use more disk space, since
> they are less costly and optimize the data warehousing.

Sorry, but quoting Stonebraker (who has a *financial* interest in his
statement), is pointless.  Similarly, you can't directly compare his
concepts to your case.

IMHO, you're trying to get buy-in.  In this group, unless you have a
patch that proves something, you're generally out of luck.  My
suggestion is to start coding.  You will find, as I did, that DSM is a
better solution.

-- 
Jonah H. Harris, Sr. Software Architect | phone: 732.331.1324
EnterpriseDB Corporation                | fax: 732.331.1301
499 Thornall Street, 2nd Floor          | jonah.harris@enterprisedb.com
Edison, NJ 08837                        | http://www.enterprisedb.com/

Ühel kenal päeval, L, 2007-10-13 kell 17:44, kirjutas Gokulakannan
Somasundaram:
> Hi,
>       I went through this article and it was good. Please have a look
> at it.
> 
> http://www.databasecolumn.com/2007/09/one-size-fits-all.html
> 
> This article was written by Michael Stonebraker, considered to be the
> founder of our database. He has mentioned that the DBMS designed in
> 1970s haven't changed according to the change that has happened in
> Hardware landscape. 

What has happened in reality, is that the speed difference between CPU,
RAM and disk speeds has _increased_ tremendously, which makes it even
more important to _decrease_ the size of stored data if you want good
performance

> The Vertica database(Monet is a open source version with the same
> principle) makes use of the very same principle. Use more disk space,
> since they are less costly and optimize the data warehousing. 

MonetDB is not about "using more disk to get better performance", but
about reducing the need to read unused data and increasing the speed by
that.

There is also a MonetDB/X100 project, which tries to make MonetOD
order(s) of magnitude faster by doing in-page compression in order to
get even more performance, see:

http://homepages.cwi.nl/~boncz/x100.html

http://www.cwi.nl/themes/ins1/publications/docs/ZuBoNeHe:DEBULL:05.pdf

> Even otherwise we are recommending Indexes with snapshot as an option.
> We are not replacing the current index scheme. So if someone feels
> that his database should run on lesser disk space, let them create the
> normal index. If he feels he can afford to have more redundant disk
> space, then he can create indexes with snapshots. We are reducing
> random I/Os at the cost of extra disk space. So definitely that's a
> good. But tech folks like us can better decide on something based on
> experiments, as Tom has pointed out. So let's see whether Indexes with
> snapshot is worth the trade-off in space. 

Agreed. 

------------
Hannu

A Dissabte 13 Octubre 2007, Gokulakannan Somasundaram va escriure:
> Even otherwise we are recommending Indexes with snapshot as an option. We
> are not replacing the current index scheme. So if someone feels that his
> database should run on lesser disk space, let them create the normal index.
> If he feels he can afford to have more redundant disk space, then he can
> create indexes with snapshots. We are reducing random I/Os at the cost of
> extra disk space. So definitely that's a good. But tech folks like us can
> better decide on something based on experiments, as Tom has pointed out. So
> let's see whether Indexes with snapshot is worth the trade-off in space.
>

There's also LucidDB [1], another open souce column based data base. But if 
you look at the features section in their web page, you'll see they use 
page-level multi-versioning. So they are avoiding the need for storing 
snapshot information for each tuple, I think that has to be kept in mind. 

I'd really like that PostgreSQL could gain some features ala Column Based 
databases, so the administrator could choose how he wants to use the 
database, but I don't think we'll be able to compete with them if they store 
snapshot informatin per page, and we're storing it per tuple, for example. So 
any step in this directoy will probably mean understanding the decisions 
they've made in their architectures.

[1] http://www.luciddb.org/

-- 
Albert Cervera i Areny
http://www.NaN-tic.com

"Gokulakannan Somasundaram" <gokul007@gmail.com> writes:

> So Indexes with snapshots will be degrading the performance only for deletes
> and only those updates, which are updating the index tuple.

Deletes never update indexes in Postgres. Increasing the size of the index
would affect vacuum, inserts, and index accesses.

--  Gregory Stark EnterpriseDB          http://www.enterprisedb.com

On 10/14/07, Gokulakannan Somasundaram <gokul007@gmail.com> wrote:

> http://www.databasecolumn.com/2007/09/one-size-fits-all.html

> > > The Vertica database(Monet is a open source version with the same
> > > principle) makes use of the very same principle. Use more disk space,
> > > since they are less costly and optimize the data warehousing.

> What i  meant there was, it has duplicated storage of certain columns of the
> table. A table with more than one projection always needs more space, than a
> table with just one projection. By doing this they are reducing the number
> of disk operations. If they are duplicating columns of data to avoid reading
> un-necessary information, we are duplicating the snapshot information to
> avoid going to the table.

Was this about Vertica or MonetDB?  I saw that article a while ago,
and I didn't see anything that suggested Vertica duplicated data, just
that it organized it differently on disk.  What are you seeing as
being duplicated?

(This is orthogonal to the current thread; I'm just curious.)

Hi,
  I think i have a initial Implementation. It has some bugs and i am working on fixing it. But to show the advantages, I want to show the number of Logical I/Os on the screen. In order to show that, i tried enabling the log_statement option in PostgreSQL.conf. But it shows only the physical reads. What i wanted was a Logical reads count( No. of  ReadBuffer calls, which is stored in ReadBufferCount variable). So i have added this stats to the bufmgr.c(function is BufferUsage, i suppose) to show Logical Reads and Physical Reads. Is this a acceptable change?
  I thought logical read count would be helpful, even for SQL tuning. Since if someone wants to tune the SQL on a test system, things might get cached and he wouldn't know how much I/O his SQL is potentially capable of. May be we can add a statistic to show how many of those ReadBuffers are pinned Buffers.

Expecting your comments.

Thanks,
Gokul.

On Sat, Oct 20, 2007 at 09:24:07AM +0530, Gokulakannan Somasundaram wrote:
> Hi,
>   I think i have a initial Implementation. It has some bugs and i am working
> on fixing it. But to show the advantages, I want to show the number of
> Logical I/Os on the screen. In order to show that, i tried enabling the
> log_statement option in PostgreSQL.conf. But it shows only the physical
> reads. What i wanted was a Logical reads count( No. of  ReadBuffer calls,
> which is stored in ReadBufferCount variable). So i have added this stats to
> the bufmgr.c(function is BufferUsage, i suppose) to show Logical Reads and
> Physical Reads. Is this a acceptable change?

I'm not sure if the number of logical reads is really a useful
measurement. I can imagine there are places that deliberatly read the
block "logically" a few times but drop the pin in between to allow
others access. This will skew your results as in actual usage only the
first is likely to generate a real I/O.

If your problem is cache it seems to me you should test with a table
larger than your shared buffers and perhaps even larger than your total
memory, since this is the case we're actually interested in.

Have a ncie day,
--
Martijn van Oosterhout   <kleptog@svana.org>   http://svana.org/kleptog/
> From each according to his ability. To each according to his ability to litigate.

Hi Hannu,

On 10/14/07 12:58 AM, "Hannu Krosing" <hannu@skype.net> wrote:

> What has happened in reality, is that the speed difference between CPU,
> RAM and disk speeds has _increased_ tremendously

Yes.

> which makes it even
> more important to _decrease_ the size of stored data if you want good
> performance

Or bring the cpu processing closer to the data it's using (or both).

By default, the trend you mention first will continue in an unending way -
the consequence is that the "distance" between a processor and it's target
data will continue to increase ad-infinitum.

By contrast, you can only decrease the data volume so much - so in the end
you'll be left with the same problem - the data needs to be closer to the
processing.  This is the essence of parallel / shared nothing architecture.

Note that we've done this at Greenplum.  We're also implementing a DSM-like
capability and are investigating a couple of different hybrid row / column
store approaches.

Bitmap index with index-only access does provide nearly all of the
advantages of a column store from a speed standpoint BTW.  Even though
Vertica is touting speed advantages - our parallel engine plus bitmap index
will crush them in benchmarks when they show up with real code.

Meanwhile they're moving on to new ideas - I kid you not "Horizontica" is
Dr. Stonebraker's new idea :-)

So - bottom line - some ideas from column store make sense, but it's not a
cure-all.
> There is also a MonetDB/X100 project, which tries to make MonetOD
> order(s) of magnitude faster by doing in-page compression in order to
> get even more performance, see:

Actually, the majority of the points made by the MonetDB team involve
decreasing the abstractions in the processing path to improve the IPC
(instructions per clock) efficiency of the executor.

We are also planning to do this by operating on data in vectors of projected
rows in the executor, which will increase the IPC by reducing I-cache misses
and improving D-cache locality.  Tight loops will make a much bigger
difference when long runs of data are the target operands.

- Luke 

Hi,
     I have tested with makeing this change and it is showing useful readings. The point of introducing the indexes with snapshot is that it should reduce the number of logical I/Os.(It may be from memory / from hard disk).  Logical I/Os are potential Physical I/Os.

Even, if the statistic doesn't become useful in some cases, we can safely ignore it.
I will submit my initial patch today.

Thanks,
Gokul.

Ühel kenal päeval, L, 2007-10-20 kell 10:19, kirjutas Luke Lonergan:
> Hi Hannu,
> 
> On 10/14/07 12:58 AM, "Hannu Krosing" <hannu@skype.net> wrote:
> 
> > What has happened in reality, is that the speed difference between CPU,
> > RAM and disk speeds has _increased_ tremendously
> 
> Yes.
> 
> > which makes it even
> > more important to _decrease_ the size of stored data if you want good
> > performance
> 
> Or bring the cpu processing closer to the data it's using (or both).
> 
> By default, the trend you mention first will continue in an unending way -
> the consequence is that the "distance" between a processor and it's target
> data will continue to increase ad-infinitum.

the emergence of solid-state (flash) disks may help a little here, but
in general it is true.

> By contrast, you can only decrease the data volume so much - so in the end
> you'll be left with the same problem - the data needs to be closer to the
> processing.  This is the essence of parallel / shared nothing architecture.
> 
> Note that we've done this at Greenplum.  We're also implementing a DSM-like
> capability and are investigating a couple of different hybrid row / column
> store approaches.

Have you tried moving the whole visibility part of tuples out to a
separate heap ?

Especially in OLAP/ETL scenarios the distribution of tuples loaded in
one transaction should be very good for visibility-info compression.

I'd suspect that you could crush hundreds of pages worth of visibility
into single RLE encoding unit (xmin=N, xmax=no_yet, start_ctid = X,
end_ctid=Y), and it will stay in L1 cache most of the time you process
the corresponding relation. and the relation itself will be smaller, and
index-only (actually index-only + lookup inside L1 cache) access can
happen, and so on .

OTOH, if you load it in millions of small transactions, you can run
VACUUM FREEZE _on_ the visibility heap only, which will make all
visibility infoe look similar and thus RLE-compressable and again make
it fit in L1 cache, if you dont have lots of failed loads interleaved
with successful ones.

> Bitmap index with index-only access does provide nearly all of the
> advantages of a column store from a speed standpoint BTW.  Even though
> Vertica is touting speed advantages - our parallel engine plus bitmap index
> will crush them in benchmarks when they show up with real code.
> 
> Meanwhile they're moving on to new ideas - I kid you not "Horizontica" is
> Dr. Stonebraker's new idea :-)

Sounds like a result of a marketroid brainstorming session :P

> So - bottom line - some ideas from column store make sense, but it's not a
> cure-all.
>  
> > There is also a MonetDB/X100 project, which tries to make MonetOD
> > order(s) of magnitude faster by doing in-page compression in order to
> > get even more performance, see:
> 
> Actually, the majority of the points made by the MonetDB team involve
> decreasing the abstractions in the processing path to improve the IPC
> (instructions per clock) efficiency of the executor.

The X100 part was about doing in-page compression, so the efficiency of
disk to L1 cache pathway would increase. so for 1/2 compression the CPU
would get twice the data threoughput.

> We are also planning to do this by operating on data in vectors of projected
> rows in the executor, which will increase the IPC by reducing I-cache misses
> and improving D-cache locality.  Tight loops will make a much bigger
> difference when long runs of data are the target operands.
> 
> - Luke 
> 
> 
> 
> ---------------------------(end of broadcast)---------------------------
> TIP 7: You can help support the PostgreSQL project by donating at
> 
>                 http://www.postgresql.org/about/donate

Gokulakannan Somasundaram wrote:
>     I would like to present the first patch. It currently has the following
> restrictions
> a) It does not support any functional indexes.
> b) It supports queries like select count(1) from table where (restrictions
> from indexed columns), but it does not support select count(1) from table.

An interesting question is how to represent tuples coming from the index
in the executor. I see that you didn't address that at all, because you
only support "COUNT(1)", and not things like "SELECT column FROM table
WHERE id = ?" where you actually return datums from the index. But
that's something that we have to think about in the DSM approach as well.

One solution is to form a heap tuple, using the datums from the index,
with the attributes that are not used in the query replaced with NULLs.
That seems simple, but I don't think it'll work with expression indexes,
when you do something like "SELECT length(column) FROM table WHERE id =
?", and there's an index on (id, length(column)).

> I have also enabled the display of Logical Reads. In order to see that, set
> log_statement_stats on.

You should start benchmarking, to verify that you're really getting the
kind of speed up you're looking for, before you spend any more effort on
that. Reduction in logical reads alone isn't enough. Remember that for a
big change like that, the gain has to be big as well.

As a first test, I'd like to see results from SELECTs on different sized
tables. On tables that fit in cache, and on tables that don't. Tables
large enough that the index doesn't fit in cache. And as a special case,
on a table just the right size that a normal index fits in cache, but a
thick one doesn't.

--
  Heikki Linnakangas
  EnterpriseDB   http://www.enterprisedb.com

Please keep the list cc'd.

Gokulakannan Somasundaram wrote:
> On 10/23/07, Heikki Linnakangas <heikki@enterprisedb.com> wrote:
>> Gokulakannan Somasundaram wrote:
>> I have also enabled the display of Logical Reads. In order to see that,
>> set
>>> log_statement_stats on.
>> You should start benchmarking, to verify that you're really getting the
>> kind of speed up you're looking for, before you spend any more effort on
>> that. Reduction in logical reads alone isn't enough. Remember that for a
>> big change like that, the gain has to be big as well.
> 
> I have done the benchmark. I have done the benchmark with Logical reads, as
> they turn out to be potential physical reads. Try turning on the
> log_statement_stats in postgresql.conf. try firing some queries, which can
> satisfied by the index. You would see the difference.

I would see a decrease in the number of logical reads, that's all. You
need to demonstrate a real increase in throughput and/or reduction in
response times.

Note that even though you reduce the number of logical reads, with a
thick index a logical read is *more* likely to be a physical read,
because the index is larger and therefore consumes more cache.

> As a first test, I'd like to see results from SELECTs on different sized
>> tables. On tables that fit in cache, and on tables that don't. Tables
>> large enough that the index doesn't fit in cache. And as a special case,
>> on a table just the right size that a normal index fits in cache, but a
>> thick one doesn't.
> 
> I have not done a Load test. That's a good idea. Are you guys using Apache
> JMeter?

You can use whatever you want, as long as you can get the relevant
numbers out of it. contrib/pgbench is a good place to start.

DBT-2 is another test people often use for patches like this. It's quite
tedious to set up and operate, but it'll give you nice very graphs.

Make sure you control vacuums, checkpoints etc., so that you get
repeatable results.

> Also i think you might have noted that the thick indexes are not affected by
> updates, if the updated column is not in the index. I think that add on to
> one more advantage of thick indexes against DSM.

That cannot possibly work. Imagine that you have a table

ctid | id | data
-----+----+-----
(0,1)| 1  | foo
(0,2)| 1  | bar

where (0,2) is an updated version of (0,1). If you don't update the
index, there will be no index pointer to (0,2), so a regular index scan,
not an index-only scan, will not find the updated tuple.

Or did you mean that the index is not updated on HOT updates? That's an
interesting observation. We could do index-only scans with the DSM as
well, even if there's HOT updates, if we define the bit in the bitmap to
mean "all tuples in this page are visible to everyone, or there's only
HOT updates". That works, because an index-only-scan doesn't access any
of the updated columns. It probably isn't worth it, though. Seems like a
pretty narrow use case, and makes it more complicated.

--  Heikki Linnakangas EnterpriseDB   http://www.enterprisedb.com

Thanks,
Gokul,
CertoSQL Project,
Allied Solution Groups.
( www.alliedgroups.com)

Gokulakannan Somasundaram wrote:
> Say, with a normal index, you need to goto the table for checking the
> snapshot. So you would be loading both the index pages + table pages, in
> order to satisfy a certain operations. Whereas in thick index you occupy 16
> bytes per tuple more in order to avoid going to the table. So memory
> management is again better. But i can run the load test, if that's
> required.

Yes, performance testing is required for any performance-related patch.

Remember that you're competing against DSM. We're going to want some
kind of a DSM anyway because it allows skipping unmodified parts of the
heap in vacuum.

>  Even when all the tuples are in memory, index only scans are
> almost 40-60% faster than the index scans with thin indexes.

Have you actually benchmarked that? What kind of query was that? I don't
believe for a second that fetching the heap tuple when the page is in
memory accounts for 40-60% of the overhead of regular index scans.

> What do you thick about not maintaining pins in case of thick indexes?

Seems irrelevant. Keeping a page pinned is cheap.

BTW, another issue you'll have to tackle, that a DSM-based patch will
have to solve as well, is how to return tuples from an index. In b-tree,
we scan pages page at a time, keeping a list of all tids that match the
scanquals in BTScanOpaque. If we need to return not only the tids of the
matching tuples, but the tuples as well, where do we store them? You
could make a palloc'd copy of them all, but that seems quite expensive.

--  Heikki Linnakangas EnterpriseDB   http://www.enterprisedb.com

Thanks,
Gokul.
CertoSQL Project.
Allied Solution Groups.
( www.alliedgroups.com)

Ühel kenal päeval, T, 2007-10-23 kell 13:04, kirjutas Heikki
Linnakangas:
> Gokulakannan Somasundaram wrote:
> > Say, with a normal index, you need to goto the table for checking the
> > snapshot. So you would be loading both the index pages + table pages, in
> > order to satisfy a certain operations. Whereas in thick index you occupy 16
> > bytes per tuple more in order to avoid going to the table. So memory
> > management is again better. But i can run the load test, if that's
> > required.
> 
> Yes, performance testing is required for any performance-related patch.
> 
> Remember that you're competing against DSM. We're going to want some
> kind of a DSM anyway because it allows skipping unmodified parts of the
> heap in vacuum.

I would suggest that you use just an additional heap with decoupled
visibility fields as DSM.

For a large number of usage scenarios this will be highly compressible
and will mostly stay in processor caches .

You can start slow, and have the info duplicated in both main heap and 
visibility heap (aka DSM).

There are several advantages to keeping a separate visibility heap:

1) it is usually higly compressible, at least you can throw away
cmin/cmax quite soon, usually also FREEZE and RLE encode the rest.

2) faster access, more tightly packed data pages.

3) index-only scans

4) superfast VACUUM FREEZE 

5) makes VACUUM faster even for worst cases (interleaving live and dead
tuples)

6) any index scan will be faster due to fetching only visible rows from
main heap.

> >  Even when all the tuples are in memory, index only scans are
> > almost 40-60% faster than the index scans with thin indexes.
> 
> Have you actually benchmarked that? What kind of query was that? I don't
> believe for a second that fetching the heap tuple when the page is in
> memory accounts for 40-60% of the overhead of regular index scans.

It depends heavily on the type of memory (postgresql page or disk cache)
it is in.

I remember doing Slony sobscribes in early days, and the speed
difference on loading a table with active PK index was several times,
depending on shared_buffers setting. 

That was for a table, where both heap and index did fit in the 2G memory
which was available, the difference being only shuffling the pages
between postgresql buffer and linux system cache or not.

> BTW, another issue you'll have to tackle, that a DSM-based patch will
> have to solve as well, is how to return tuples from an index. In b-tree,
> we scan pages page at a time, keeping a list of all tids that match the
> scanquals in BTScanOpaque. If we need to return not only the tids of the
> matching tuples, but the tuples as well, where do we store them? You
> could make a palloc'd copy of them all, but that seems quite expensive.

Have you considered returning them as "already visibility-checked pages"
similar to what views or set-returning functions return ?

-------------------
Hannu

--
Thanks,
Gokul.
CertoSQL Project,
Allied Solution Groups.
(www.alliedgroups.com )

Hannu Krosing wrote:
> I would suggest that you use just an additional heap with decoupled
> visibility fields as DSM.

Yeah, I remember you've suggested that before, and I haven't responded
this far. The problems I see with that approach are:

1) How do you know which visibility info corresponds which heap tuple?
You'd need to have a pointer from the visibility info to the heap tuple,
and from the heap tuple to the visibility info. Which increases the
total (uncompressed) storage size.

2) If the visibility info / heap ordering isn't the same, seqscans need
to do random I/O.

3) If you need to do regular index scans, you're going to have to access
the index, the heap and the visibility info separately, and in that
order. That sounds expensive.

4) It's a big and complex change.

The significance of 2 and 3 depends a lot on how much of the visibility
information is in cache.

> For a large number of usage scenarios this will be highly compressible
> and will mostly stay in processor caches .

This seems to be where the potential gains are coming from in this
scheme. It boils down to how much compression you can do, and how
expensive it is to access the information in compressed form.

> 1) it is usually higly compressible, at least you can throw away
> cmin/cmax quite soon, usually also FREEZE and RLE encode the rest.

If you RLE compress the data, you'll need to figure out what to do when
you need update a field and it doesn't compress as well anymore. You
might have to move things around pages, so you'll have to update any
pointers to that information atomically.

> 2) faster access, more tightly packed data pages.

But you do need to access the visibility information as well, at least
on tuples that match the query.

> 5) makes VACUUM faster even for worst cases (interleaving live and dead
> tuples)

Does it? You still need to go to the heap pages to actually remove the
dead tuples. I suppose you could skip that and do it the first time you
access the page, like we do pruning with HOT.

> 6) any index scan will be faster due to fetching only visible rows from
> main heap.

Assuming the visibility information is already in cache, and that
there's enough non-visible tuples for that to matter.

>> BTW, another issue you'll have to tackle, that a DSM-based patch will
>> have to solve as well, is how to return tuples from an index. In b-tree,
>> we scan pages page at a time, keeping a list of all tids that match the
>> scanquals in BTScanOpaque. If we need to return not only the tids of the
>> matching tuples, but the tuples as well, where do we store them? You
>> could make a palloc'd copy of them all, but that seems quite expensive.
> 
> Have you considered returning them as "already visibility-checked pages"
> similar to what views or set-returning functions return ?

Sorry, I don't understand what you mean by that.

--  Heikki Linnakangas EnterpriseDB   http://www.enterprisedb.com

Ühel kenal päeval, T, 2007-10-23 kell 18:36, kirjutas Gokulakannan
Somasundaram:

>         
>         There are several advantages to keeping a separate visibility
>         heap:
>         
>         1) it is usually higly compressible, at least you can throw
>         away 
>         cmin/cmax quite soon, usually also FREEZE and RLE encode the
>         rest.
>         
>         2) faster access, more tightly packed data pages.
>         
>         3) index-only scans
>         
>         4) superfast VACUUM FREEZE
>         
>         5) makes VACUUM faster even for worst cases (interleaving live
>         and dead 
>         tuples)
>         
>         6) any index scan will be faster due to fetching only visible
>         rows from
>         main heap.
> 
> if you have to store the visibility fields of all the tuples of each
> table, then you may not be able to accomodate in the cache. Say if a
> table is  of 1 million rows, we would need 22 MB of  visibility
> space(since visibility info takes 16 bytes. I think if we have to link
> it with say tuple-id(6 Bytes). 

You can keep the visibility info small, by first dropping cmin/cmax and
then FREEZ'ing the tuples (setting xmin to special value), after that
you can replace a lot of visibility info tuples with single RLE encoded
tuple, which simply states, that tuples N:A to M:B are visible.

If that 1 million row table is mostly static, the static parts will soon
have (al lot) less than 1 bit in visibility heap.

For example, after vacuum there will be just one visibility info which
say that whole table is visible.

I envision HOT-like on-the-fly VACUUM FREEZE manipulations of visibility
info so it won't grow very big at all.

> I think we may need to link it with indexes with one more id. i am not
> counting that now). 

why ?

we will keep visibility info for ctids (PAGE:NR) and if we need to see,
if any ctid pointe from index points to a visible tuple we check it
based on that ctid.

>  If we have  10 tables, then we will have 220 MB.  Keeping them pinned
> in memory may not be advisable in some circumstances. 

no no! no pinning, the "mostly in cache" will happen automatically (and
I mean mostly in processors _internal_ L1 or L2 cache, not just in RAM)

> If it is not going to be in memory, then that is no different from
> referring a table. But i accept that is a concept worth trying out. I
> think the advantage with thick indexes comes with the fact, that it is
> optional. If we can make this also as optional, that would be better. 
> But if we are going to suggest it as a replacement of DSM, then it
> loses the advantage of being small. 

I agree that a single-purpose DSM can be made smaller than multi-purpose
visibility heap.


------------------
Hannu

Ühel kenal päeval, T, 2007-10-23 kell 14:16, kirjutas Heikki
Linnakangas:
> Hannu Krosing wrote:
> > I would suggest that you use just an additional heap with decoupled
> > visibility fields as DSM.
> 
> Yeah, I remember you've suggested that before, and I haven't responded
> this far. The problems I see with that approach are:
>
> 1) How do you know which visibility info corresponds which heap tuple?
> You'd need to have a pointer from the visibility info to the heap tuple,
> and from the heap tuple to the visibility info. Which increases the
> total (uncompressed) storage size.

only the first , and not a pointer mostly , but just a fast lookup
function IsVisible(ctid, snapshot)

> 2) If the visibility info / heap ordering isn't the same, seqscans need
> to do random I/O.

Visibility should be a tree-like structure, with
wholly-(visible/invisible) inner pages removed, stored in seqscan order.

seqscans will get a few (or few hundred) pages worth of sequential info
and then fetch only visible tuples from heap, then repeat

> 3) If you need to do regular index scans, you're going to have to access
> the index, the heap and the visibility info separately, and in that
> order. That sounds expensive.

I'd do it in order 1) index 2) visibility 3) heap

The main assumption is, that visibility info is highly compressed -->
stays in processor cache most of the time --> virtually free to access

> 4) It's a big and complex change.

We can go in smaller steps, first adding it as a DSM replacement and as
a complement to visibility info in heap and only when we are sure that
we have gotten it right (by checking that there is always valid info in
visibility heap) will we move other internal stuff to using it, one by
one.


> The significance of 2 and 3 depends a lot on how much of the visibility
> information is in cache.
> 
> > For a large number of usage scenarios this will be highly compressible
> > and will mostly stay in processor caches .
> 
> This seems to be where the potential gains are coming from in this
> scheme. It boils down to how much compression you can do, and how
> expensive it is to access the information in compressed form.

Hopefully much cheaper than in uncompressed form, the best case being
when the table is static and you do an in-L1-cache lookup on VizHeap
root page

> > 1) it is usually higly compressible, at least you can throw away
> > cmin/cmax quite soon, usually also FREEZE and RLE encode the rest.
> 
> If you RLE compress the data, you'll need to figure out what to do when
> you need update a field and it doesn't compress as well anymore. You
> might have to move things around pages, so you'll have to update any
> pointers to that information atomically.

I envision each RLE range to have an exception list, which can be
appended to atomically with no locks. only when the list is long enough
will there be locking and rearranging.

> > 2) faster access, more tightly packed data pages.
> 
> But you do need to access the visibility information as well, at least
> on tuples that match the query.

usually you even access visbility first, then data (if visible).

> > 5) makes VACUUM faster even for worst cases (interleaving live and dead
> > tuples)
> 
> Does it? You still need to go to the heap pages to actually remove the
> dead tuples. I suppose you could skip that and do it the first time you
> access the page, like we do pruning with HOT.

Possibly for heap. But not for indexes which have to be cleaned up
during vacuum. At least I can't see how we could do it later.

OTOH, we can have very fast VACUUM FREEZE ONLY command, which will run
on just visibility heap and FREEZE all visible-to-all tuples and
compress them. It should also mark all aborted or
deleted-and-invisible-to-all to a separate value so they too compress
better.

> > 6) any index scan will be faster due to fetching only visible rows from
> > main heap.
> 
> Assuming the visibility information is already in cache, and that
> there's enough non-visible tuples for that to matter.

If there is not enough, then the visibility info is likely very highly
compressible and thus cheap to access.

----------------
Hannu

--
Thanks,
Gokul.
CertoSQL Project,
Allied Solution Groups.
(www.alliedgroups.com )