Thread: [HACKERS] [WIP]Vertical Clustered Index (columnar store extension)

On Fri, Dec 30, 2016 at 02:55:39PM +1100, Haribabu Kommi wrote:
> 
> Hi All,
> 
> Fujitsu was interested in developing a columnar storage extension with minimal
> changes the server backend.
> 
> The columnar store is implemented as an extension using index access methods.
> This can be easily enhanced with pluggable storage methods once they are
> available.

Have you see this post from 2015:
https://www.postgresql.org/message-id/20150831225328.GM2912%40alvherre.pgsql

--  Bruce Momjian  <bruce@momjian.us>        http://momjian.us EnterpriseDB
http://enterprisedb.com

+ As you are, so once was I.  As I am, so you will be. +
+                      Ancient Roman grave inscription +

On 12/29/16 9:55 PM, Haribabu Kommi wrote:
> The tuples which don't have multiple copies or frozen data will be moved
> from WOS to ROS periodically by the background worker process or autovauum
> process. Every column data is stored separately in it's relation file. There
> is no transaction information is present in ROS. The data in ROS can be
> referred with tuple ID.

Would updates be handled via the delete mechanism you described then?

> In this approach, the column data is present in both heap and columnar
> storage.

ISTM one of the biggest reasons to prefer a column store over heap is to 
ditch the 24 byte overhead, so I'm not sure how much of a win this is.

Another complication is that one of the big advantages of a CSTORE is 
allowing analysis to be done efficiently on a column-by-column (as 
opposed to row-by-row) basis. Does your patch by chance provide that?

Generally speaking, I do think the idea of adding support for this as an 
"index" is a really good starting point, since that part of the system 
is pluggable. It might be better to target getting only what needs to be 
in core into core to begin with, allowing the other code to remain an 
extension for now. I think there's a lot of things that will be 
discovered as we start moving into column stores, and it'd be very 
unfortunate to accidentally paint the core code into a corner somewhere.

As a side note, it's possible to get a lot of the benefits of a column 
store by using arrays. I've done some experiments with that and got an 
80-90% space reduction, and most queries saw improved performance as 
well (there were a few cases that weren't better). The biggest advantage 
to this approach is people could start using it today, on any recent 
version of Postgres. That would be a great way to gain knowledge on what 
users would want to see in a column store, something else I suspect we 
need. It would also be far less code than what you or Alvaro are 
proposing. When it comes to large changes that don't have crystal-clear 
requirements, I think that's really important.
-- 
Jim Nasby, Data Architect, Blue Treble Consulting, Austin TX
Experts in Analytics, Data Architecture and PostgreSQL
Data in Trouble? Get it in Treble! http://BlueTreble.com
855-TREBLE2 (855-873-2532)

On 12/29/16 10:55 PM, Haribabu Kommi wrote:
> Fujitsu was interested in developing a columnar storage extension with
> minimal
> changes the server backend.
> 
> The columnar store is implemented as an extension using index access
> methods.
> This can be easily enhanced with pluggable storage methods once they are
> available.
> 
> A new index method (VCI) is added to create columnar index on the table.

I'm confused.  You say that you are adding an index access method, for
which we have a defined extension mechanism, but the code doesn't do
that.  Instead, it sprinkles a bunch of hooks through the table access
code.  So you are really adding ways to add alternatives to heap
storage, except we have no way to know whether these hooks have been
designed with any kind of generality in mind.  So is it an index access
method or a table access method?

Either way, you shouldn't need a new relkind.  Note that all indexes
have the same relkind, even if they use different access methods.

I think there are two ways to integrate column storage into PostgreSQL:
One is to use the FDW interface.  That has been done before, see
cstore_fdw.  The other is to define a storage manager extension
interface.  That has been tried but has not been completed yet.  Adding
a bunch of custom hooks all over the place seems worse than both of those.

-- 
Peter Eisentraut              http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

On 1/16/17 10:09 PM, Haribabu Kommi wrote:
> Yes, that' correct. Currently with this approach, it is not possible to
> ditch the
> heap completely. This approach is useful for the cases, where the user wants
> to store only some columns as part of clustered index.

Ahh, that's unfortunate. Billion row+ tables are becoming rather common, 
and that 24GB of overhead starts becoming very painful. It's actually a 
lot worse considering there will be at least one index on the table, so 
100GB+ of overhead isn't that uncommon.

>     Another complication is that one of the big advantages of a CSTORE
>     is allowing analysis to be done efficiently on a column-by-column
>     (as opposed to row-by-row) basis. Does your patch by chance provide
>     that?
>
> Not the base patch that I shared. But the further patches provides the
> data access
> column-by-column basis using the custom plan methods.

Great, that's something else that a column store really needs to be 
successful. Something else I suspect is necessary is a faster/better way 
to eliminate chunks of rows from scans.

Just as an example, with my simple array-based approach, you can store a 
range type along with each array that contains the min and max values 
for the array. That means any query that wants values between 50 and 100 
can include a clause that filters on range types that overlap with 
[50,100]. That can be indexed very efficiently and is fast to run checks 
against.

>     Generally speaking, I do think the idea of adding support for this
>     as an "index" is a really good starting point, since that part of

... as discussed elsewhere in the thread, adding a bunch of hooks is 
probably not a good way to do this. :/

>     That would be a great way to gain knowledge on what users would want
>     to see in a column store, something else I suspect we need. It would
>     also be far less code than what you or Alvaro are proposing. When it
>     comes to large changes that don't have crystal-clear requirements, I
>     think that's really important.
>
> The  main use case of this patch is to support mixed load environments,
> where both OLTP and OLAP queries are possible. The advantage of
> proposed patch design is, providing good performance to OLAP queries
> without affecting OLTP.

Yeah, that's a big part of what I was envisioning with my array-based 
approach. In simple terms, there would be a regular row-based table, and 
an array-based table, with a view that allows seamless querying into 
both (re-presenting the array-storage on a per-row basis). There would 
be a periodic process that moves entire sets of rows from the row 
storage into the array storage.

If you updated or deleted a row that was part of an array, the contents 
of the entire array could be moved back into row-based storage. After a 
period of time, rows would get moved back into array storage. Or the 
array could be modified in place, but you need to be very careful about 
bloating the array storage if you do that.

The big missing piece here is getting the planner to intelligently 
handle a mixed row/column store. As I mentioned, you can easily add 
range type fields to greatly increase performance, but they won't do any 
good unless the appropriate filters get added. It's not THAT hard to do 
that by hand, but it'd be great if there was a more automated method. 
Such a method might also be very useful for transforming expressions 
like date_part('quarter', ...) into something that could use existing 
indexes.
-- 
Jim Nasby, Data Architect, Blue Treble Consulting, Austin TX
Experts in Analytics, Data Architecture and PostgreSQL
Data in Trouble? Get it in Treble! http://BlueTreble.com
855-TREBLE2 (855-873-2532)

On Fri, Dec 30, 2016 at 12:55 PM, Haribabu Kommi
<kommi.haribabu@gmail.com> wrote:
> Any Comments on the approach?

I have moved this patch to CF 2017-03.
-- 
Michael

On 30.12.2016 06:55, Haribabu Kommi wrote:

Hi All,

Fujitsu was interested in developing a columnar storage extension with minimal

changes the server backend.

We  in PostgresPRO are also very interested in developing vertical storage (VS) for Postgres.
And after considering many alternatives, we came to the conclusion that approach based on representing columnar store as access method (index)
is the most promising one.

It allows to:
1. Implement VS as extension without affecting Postgres core.
2. Have both ROS and WOS.
3. Create multiple projections (as in Vertica).
4. Optimize insert speed by support batch inserts and use flexible recovery model for VS.

So it is very similar with your approach. But there are few differences:

1. Our intention is to completely eliminate changes in Postgres core.

You wrote:

Yes, it is a mix of both index and table access methods. The current design

of Vertical clustered index needs both access methods, because of this reason

we used both access methods.

But I still do not completely understand why it is not possible to use VS in index only scans without any changes and standard Postgres executor?
Why it is not possible to rely on standard rules of applying indexes in Postgres optimizer based on costs provided by our AM implementation?


2. You are accessing VS pages through Postgres buffer manager. It certainly have a lot of advantages. First of all it significantly simplifies implementation of VS and allows to reuse Postgres cache and lock managers.
But is all leads to some limitation:
- For VS it is preferable to have larger pages (in Vertica size of page can be several megabytes).
- VS is optimized for sequential access, so caching pages in buffer manager is no needed and can only cause leaching of other useful pages from cache.
- It makes it not possible to implement in-memory version of VS.
- Access to buffer manager adds extra synchronization overhead which becomes noticeable at MPP systems.

So I wonder if you have considered approach with VS specific implementation of storage layer?

3. To take all advantages of vertical model, we should provide vector execution.
Without it columnar store can only reduce amount of fetched data by selective fetch of accessed columns and better compression of them.
But this is what existed cstore_fdw extension for Postgres also does.

We are going to use executor hooks or custom nodes to implement vector operations for some nodes (filter, grand aggregate, aggregation with group by,...).
Something similar with  https://github.com/citusdata/postgres_vectorization_test

What is your vision of optimizing executor to work with VS?

4. How do you consider adding parallelism support to VS? Should it be handled inside VS implementation? Or should we use standard Postgres parallel execution (parallel index-only scan)?

Thanks in advance,
Kosntantin

The columnar store is implemented as an extension using index access methods.

This can be easily enhanced with pluggable storage methods once they are available.

A new index method (VCI) is added to create columnar index on the table.

The following is the basic design idea of the columnar extension, 

This has the on-disk columnar representation. So, even after crash, 

the columnar format is recovered to the state when it was crashed.

To provide performance benefit for both read and write operations,

the data is stored in two formats

1) write optimized storage (WOS) 

2) read optimized storage (ROS).

This is useful for the users where there is a great chance of data modification

that is newly added instead of the old data.

WOS

====

write optimized storage is the data of all columns that are part of VCI are

stored in a row wise format. All the newly added data is stored in WOS

relation with xmin/xmax information also. If user wants to update/delete the

newly added data, it doesn't affect the performance much compared to

deleting the data from columnar storage.

The tuples which don't have multiple copies or frozen data will be moved

from WOS to ROS periodically by the background worker process or autovauum

process. Every column data is stored separately in it's relation file. There

is no transaction information is present in ROS. The data in ROS can be

referred with tuple ID.

In this approach, the column data is present in both heap and columnar

storage.

ROS

====

This is the place, where all the column data is stored in columnar format.

The data from WOS to ROS is converted by background workers continously based

on the tuple visibility check. Whenever the tuple is frozen and it gets moved

from WOS to ROS.

The Data in ROS is stored in extents. One extent contains of 262,144 rows. Because

of fixed number of records in an extent it is easy to map the heap record to the columnar

record with TID to CRID map.

Insert

=====

The insert operation is just like inserting a data into an index.

Select

=====

Because of two storage formats, during the select operation, the data in WOS

is converted into Local ROS for the statement to be executed. The conversion

cost depends upon the number of tuples present in the WOS file. This

may add some performance overhead for select statements. The life of the Local

ROS is till the end of query context.

Delete

=====

During the delete operation, whenever the data is deleted in heap at the same

time the data in WOS file is marked as deleted similar like heap. But in case

if the data is already migrated from WOS to ROS, then we will maintain some

delete vector to store the details of tuple id, transaction information and etc.

During the data read from ROS file, it is verified against delete vector and 

confirms whether the record is visible or not? All the delete vectors

data is applied to ROS periodically.

More details of internal relations and their usage is available in the README.

Still it needs more updates to explain full details of the columnar index design.

The concept of Vertical clustered index columnar extension is from Fujitsu Labs, Japan.

Following is the brief schedule of patches that are required

for a better performing columnar store.

1. Minimal server changes (new relkind "CSTORE" option)

2. Base storage patch

3. Support for moving data from WOS to ROS

4. Local ROS support

5. Custom scan support to read the data from ROS and Local ROS

6. Background worker support for data movement

7. Expression state support in VCI

8. Aggregation support in VCI 

9. Pg_dump support for the new type of relations

10. psql \d command support for CSTORE relations

11. Parallelism support

12. Compression support

13. In-memory support with dynamic shared memory

Currently I attached only patches for 1 and 2. These will provide the

basic changes that are required in PostgreSQL core to the extension

to work.

I have to rebase/rewrite the rest of the patches to the latest master, 

and share them with community. 

Any Comments on the approach?

 

Regards,

Hari Babu

Fujitsu Australia

-- 
Konstantin Knizhnik
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company 

Hi,

I wonder if it is possible to somehow benchmark your clustered index implementation.
I tried to create VCI index for lineitem table from TPC and run Q6 query.
After index creation Postgres is not using parallel execution plan any more but speed of sequential plan is not changed
and nothing in query execution plan indicates that VCI index is used:


postgres=# explain select
    sum(l_extendedprice*l_discount) as revenue
from
    lineitem_projection
where
    l_shipdate between '1996-01-01' and '1997-01-01'
    and l_discount between 0.08 and 0.1
    and l_quantity < 24;
                                                                                                                 QUERY PLAN                           
                                                                                    
-------------------------------------------------------------------------------------------------------------------------------------------------------
-------------------------------------------------------------------------------------
 Finalize Aggregate  (cost=608333.85..608333.86 rows=1 width=4)
   ->  Gather  (cost=608333.23..608333.84 rows=6 width=4)
         Workers Planned: 6
         ->  Partial Aggregate  (cost=607333.23..607333.24 rows=1 width=4)
               ->  Parallel Seq Scan on lineitem_projection  (cost=0.00..607024.83 rows=61680 width=8)
                     Filter: ((l_shipdate >= '1996-01-01'::date) AND (l_shipdate <= '1997-01-01'::date) AND (l_discount >= '0.08'::double precision) AN
D (l_discount <= '0.1'::double precision) AND (l_quantity < '24'::double precision))
(6 rows)

postgres=# select
    sum(l_extendedprice*l_discount) as revenue
from
    lineitem_projection
where
    l_shipdate between '1996-01-01' and '1997-01-01'
    and l_discount between 0.08 and 0.1
    and l_quantity < 24;
   revenue  
-------------
 6.21111e+08
(1 row)

Time: 1171.324 ms (00:01.171)

postgres=# create index vci_idx on lineitem_projection using vci(l_shipdate,l_quantity,l_extendedprice,l_discount,l_tax,l_returnflag,l_linestatus);
CREATE INDEX
Time: 4.705 ms


postgres=# explain select
    * from
    lineitem_projection
where                 
    l_shipdate between '1996-01-01' and '1997-01-01'
    and l_discount between 0.08 and 0.1
    and l_quantity < 24;
                                                                                                        QUERY PLAN                                    
                                                                  
-------------------------------------------------------------------------------------------------------------------------------------------------------
-------------------------------------------------------------------
 Seq Scan on lineitem_projection  (cost=0.00..382077.00 rows=1 width=22)
   Filter: ((l_shipdate >= '1996-01-01'::date) AND (l_shipdate <= '1997-01-01'::date) AND (l_discount >= '0.08'::double precision) AND (l_discount <= '
0.1'::double precision) AND (l_quantity < '24'::double precision))
(2 rows)

postgres=# select                                                                                                                                 
    sum(l_extendedprice*l_discount) as revenue
from
    lineitem_projection
where
    l_shipdate between '1996-01-01' and '1997-01-01'
    and l_discount between 0.08 and 0.1
    and l_quantity < 24;
  revenue  
------------
 6.2112e+08
(1 row)

Time: 4304.355 ms (00:04.304)


I wonder if there is any query which can demonstrate advantages of using VCI index?

On Fri, Feb 3, 2017 at 8:28 PM, Konstantin Knizhnik <k.knizhnik@postgrespro.ru> wrote:

On 30.12.2016 06:55, Haribabu Kommi wrote:

Hi All,

Fujitsu was interested in developing a columnar storage extension with minimal

changes the server backend.

We  in PostgresPRO are also very interested in developing vertical storage (VS) for Postgres.
And after considering many alternatives, we came to the conclusion that approach based on representing columnar store as access method (index)
is the most promising one.

It allows to:
1. Implement VS as extension without affecting Postgres core.
2. Have both ROS and WOS.
3. Create multiple projections (as in Vertica).
4. Optimize insert speed by support batch inserts and use flexible recovery model for VS.

So it is very similar with your approach. But there are few differences:

1. Our intention is to completely eliminate changes in Postgres core.

You wrote:

Yes, it is a mix of both index and table access methods. The current design

of Vertical clustered index needs both access methods, because of this reason

we used both access methods.

But I still do not completely understand why it is not possible to use VS in index only scans without any changes and standard Postgres executor?
Why it is not possible to rely on standard rules of applying indexes in Postgres optimizer based on costs provided by our AM implementation?

In our storage design, we used TID-CRID map to identify a record in heap

to columnar storage. Because of HOT update, the new data will not be inserted

into indexes, but this will give problem to the columnar storage, so we added

a hook to insert index data even if the update is HOT. 

And also we added another hook for initializing the parameters during the

execution.

Most of the other added hooks can be replaced with existing hooks and adding

some extra code.

 

2. You are accessing VS pages through Postgres buffer manager. It certainly have a lot of advantages. First of all it significantly simplifies implementation of VS and allows to reuse Postgres cache and lock managers.
But is all leads to some limitation:
- For VS it is preferable to have larger pages (in Vertica size of page can be several megabytes).
- VS is optimized for sequential access, so caching pages in buffer manager is no needed and can only cause leaching of other useful pages from cache.
- It makes it not possible to implement in-memory version of VS.
- Access to buffer manager adds extra synchronization overhead which becomes noticeable at MPP systems.

So I wonder if you have considered approach with VS specific implementation of storage layer?

Currently, we are just using the existing the PostgreSQL buffer manager 

and didn't evaluate any columnar storage specific storage implementation.

we are having some plan of evaluating dynamic shared memory.

 

3. To take all advantages of vertical model, we should provide vector execution.
Without it columnar store can only reduce amount of fetched data by selective fetch of accessed columns and better compression of them.
But this is what existed cstore_fdw extension for Postgres also does.

We are going to use executor hooks or custom nodes to implement vector operations for some nodes (filter, grand aggregate, aggregation with group by,...).
Something similar with  https://github.com/citusdata/postgres_vectorization_test

What is your vision of optimizing executor to work with VS?

Yes, we implemented similar like above by copy/paste the most of the aggregate and etc code

into the extension for providing the vector execution support.

Without this vector execution and parallelism support, there will not be much performance

benefit.

4. How do you consider adding parallelism support to VS? Should it be handled inside VS implementation? Or should we use standard Postgres parallel execution (parallel index-only scan)?

Currently we implemented our own parallelism in columnar storage with some base infrastructure

of OSS, but we are planning to change/integrate according to the OSS implementation.

Regards,

Hari Babu

Fujitsu Australia

-- 
Konstantin Knizhnik
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company 

On 2/13/17 8:59 PM, Haribabu Kommi wrote:

> The current patch that I shared doesn't contains the plan and executor
> changes to show
> the performance benefit of the clustered index. we used custom plan to
> generate the plan
> for the clustered index. Currently I am working on it to rebase it to
> current master and
> other necessary changes.
> 
> In the current state of the patch, I cannot take any performance tests,
> as it needs some
> major changes according to the latest PostgreSQL version. I have an old
> performance
> report that is took on 9.5 attached for your reference.
> 
> The current patch that is shared is to find out the best approach in
> developing a columnar
> storage in PostgreSQL, by adopting Index access methods + additional
> hooks or pluggable
> storage access methods?

While this looks like it could be a really significant performance
improvement, I think the above demonstrates that it needs a lot of work.I know this is not new to the 2017-03 CF but it
doesn'tseem enough
 
progress has been made since posting to allow it to be committed in time
for v10.

I recommend moving this patch to the 2017-07 CF.

-- 
-David
david@pgmasters.net

On 3/3/17 16:16, David Steele wrote:
> While this looks like it could be a really significant performance
> improvement, I think the above demonstrates that it needs a lot of work.
>  I know this is not new to the 2017-03 CF but it doesn't seem enough
> progress has been made since posting to allow it to be committed in time
> for v10.
> 
> I recommend moving this patch to the 2017-07 CF.

I think the patch that was in 2017-01 was given some feedback that put
the fundamental approach in question, which the author appeared to agree
with.  So I don't know why this patch appeared in this CF at all.

-- 
Peter Eisentraut              http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services

On 3/4/17 8:33 AM, Peter Eisentraut wrote:
> On 3/3/17 16:16, David Steele wrote:
>> While this looks like it could be a really significant performance
>> improvement, I think the above demonstrates that it needs a lot of work.
>>  I know this is not new to the 2017-03 CF but it doesn't seem enough
>> progress has been made since posting to allow it to be committed in time
>> for v10.
>>
>> I recommend moving this patch to the 2017-07 CF.
> 
> I think the patch that was in 2017-01 was given some feedback that put
> the fundamental approach in question, which the author appeared to agree
> with.  So I don't know why this patch appeared in this CF at all.

Then it sounds like it should be marked RWF.  Haribabu can resubmit when
there's a new candidate patch.

-- 
-David
david@pgmasters.net