Thread: Reducing planning time on tables with many indexes

David Geier <geidav.pg@gmail.com> writes:
> We tracked down the root cause of this slowdown to lock contention in
> 'get_relation_info()'. The index lock of every single index of every single
> table used in that query is acquired. We attempted a fix by pre-filtering
> out all indexes that anyways cannot be used with a certain query, without
> taking the index locks (credits to Luc Vlaming for idea and
> implementation). The patch does so by caching the columns present in every
> index, inside 'struct Relation', similarly to 'rd_indexlist'.

I wonder how much thought you gave to the costs imposed by that extra
cache space.  We have a lot of users who moan about relcache bloat
already.  But more to the point, I do not buy the assumption that
an index's set of columns is a good filter for which indexes are of
interest.  A trivial counterexample from the regression database is

regression=# explain select count(*) from tenk1;
                                         QUERY PLAN

--------------------------------------------------------------------------------
------------
 Aggregate  (cost=219.28..219.29 rows=1 width=8)
   ->  Index Only Scan using tenk1_hundred on tenk1  (cost=0.29..194.28 rows=100
00 width=0)
(2 rows)

It looks to me like the patch also makes unwarranted assumptions about
being able to discard all but the smallest index having a given set
of columns.  This would, for example, possibly lead to dropping the
index that has the most useful sort order, or that has the operator
class needed to support a specific WHERE clause.

In short, I'm not sure I buy this concept at all.  I think it might
be more useful to attack the locking overhead more directly.  I kind
of wonder why we need per-index locks at all during planning ---
I think that we already interpret AccessShareLock on the parent table
as being sufficient to block schema changes on existing indexes.

Unfortunately, as things stand today, the planner needs more than the
right to look at the indexes' schemas, because it makes physical accesses
to btree indexes to find out their tree height (and I think there are some
comparable behaviors in other AMs).  I've never particularly cared for
that implementation, and would be glad to rip out that behavior if we can
find another way.  Maybe we could persuade VACUUM or ANALYZE to store that
info in the index's pg_index row, or some such, and then the planner
could use it with no lock?

            regards, tom lane

I wrote:
> Unfortunately, as things stand today, the planner needs more than the
> right to look at the indexes' schemas, because it makes physical accesses
> to btree indexes to find out their tree height (and I think there are some
> comparable behaviors in other AMs).  I've never particularly cared for
> that implementation, and would be glad to rip out that behavior if we can
> find another way.  Maybe we could persuade VACUUM or ANALYZE to store that
> info in the index's pg_index row, or some such, and then the planner
> could use it with no lock?

A first step here could just be to postpone fetching _bt_getrootheight()
until we actually need it during cost estimation.  That would avoid the
need to do it at all for indexes that indxpath.c discards as irrelevant,
which is a decision made on considerably more information than the
proposed patch uses.

Having done that, you could look into revising plancat.c to fill the
IndexOptInfo structs from catcache entries instead of opening the
index per se.  (You'd have to also make sure that the appropriate
index locks are acquired eventually, for indexes the query does use
at runtime.  I think that's the case, but I'm not sure if anything
downstream has been optimized on the assumption the planner did it.)

This'd probably get us a large part of the way there.  Further
optimization of acquisition of tree height etc could be an
optional follow-up.

            regards, tom lane

On 27.07.22, 18:39, "Tom Lane" <tgl@sss.pgh.pa.us> wrote:

    [External Email]


    David Geier <geidav.pg@gmail.com> writes:
    > We tracked down the root cause of this slowdown to lock contention in
    > 'get_relation_info()'. The index lock of every single index of every single
    > table used in that query is acquired. We attempted a fix by pre-filtering
    > out all indexes that anyways cannot be used with a certain query, without
    > taking the index locks (credits to Luc Vlaming for idea and
    > implementation). The patch does so by caching the columns present in every
    > index, inside 'struct Relation', similarly to 'rd_indexlist'.

    I wonder how much thought you gave to the costs imposed by that extra
    cache space.  We have a lot of users who moan about relcache bloat
    already.  But more to the point, I do not buy the assumption that
    an index's set of columns is a good filter for which indexes are of
    interest.  A trivial counterexample from the regression database is

    regression=# explain select count(*) from tenk1;
                                             QUERY PLAN

    --------------------------------------------------------------------------------
    ------------
     Aggregate  (cost=219.28..219.29 rows=1 width=8)
       ->  Index Only Scan using tenk1_hundred on tenk1  (cost=0.29..194.28 rows=100
    00 width=0)
    (2 rows)

    It looks to me like the patch also makes unwarranted assumptions about
    being able to discard all but the smallest index having a given set
    of columns.  This would, for example, possibly lead to dropping the
    index that has the most useful sort order, or that has the operator
    class needed to support a specific WHERE clause.

Thanks for checking out the patch!

Just to make sure we're on the same page: we're only making this assumption if you select no fields at all.
If you select any fields at all it will check for column overlap, and if there's any overlap with any referenced field,

then the index will not be filtered out.

For producing a row count with no referenced fields it is true that it should select the truly cheapest 
index to produce the row count and there should be some Index-am callback introduced for that. 
For now it was just a quick-and-dirty solution.
Wouldn't a callback that would estimate the amount of data read be good enough though?

For sort orders the field to sort by should be listed and hence the index should not be filtered out,
or what am I missing? Likely I've missed some fields that are referenced somehow (potentially indirectly),
but that shouldn't disqualify the approach completely.

    In short, I'm not sure I buy this concept at all.  I think it might
    be more useful to attack the locking overhead more directly.  I kind
    of wonder why we need per-index locks at all during planning ---
    I think that we already interpret AccessShareLock on the parent table
    as being sufficient to block schema changes on existing indexes.

Could you elaborate as to why this approach is not good enough? To me it seems that avoiding work
ahead of time is generally useful. Or are you worried that we remove too much?

    Unfortunately, as things stand today, the planner needs more than the
    right to look at the indexes' schemas, because it makes physical accesses
    to btree indexes to find out their tree height (and I think there are some
    comparable behaviors in other AMs).  I've never particularly cared for
    that implementation, and would be glad to rip out that behavior if we can
    find another way.  Maybe we could persuade VACUUM or ANALYZE to store that
    info in the index's pg_index row, or some such, and then the planner
    could use it with no lock?

                            regards, tom lane


The thing you're touching on is specific for a btree. Not sure this generalizes to all index types that
are out there though? I could see there being some property that allows you to be "no-lock",
and then a callback that allows you to cache some generic data that can be transformed
when the indexopt info structs are filled. Is that roughly what you have in mind?

Best,
Luc

Hi Tom,

On Wed, Jul 27, 2022 at 7:15 PM Tom Lane <tgl@sss.pgh.pa.us> wrote:

I wrote:
> Unfortunately, as things stand today, the planner needs more than the
> right to look at the indexes' schemas, because it makes physical accesses
> to btree indexes to find out their tree height (and I think there are some
> comparable behaviors in other AMs).  I've never particularly cared for
> that implementation, and would be glad to rip out that behavior if we can
> find another way.  Maybe we could persuade VACUUM or ANALYZE to store that
> info in the index's pg_index row, or some such, and then the planner
> could use it with no lock?

It seems like _bt_getrootheight() first checks if the height is cached and only if it isn't it accesses index meta pages.

If the index locks are only taken for the sake of _bt_getrootheight() accessing index meta pages in case they are not cached, maybe the index locks could be taken conditionally.

However, postponing the call to where it is really needed sounds even better.

A first step here could just be to postpone fetching _bt_getrootheight()
until we actually need it during cost estimation.  That would avoid the
need to do it at all for indexes that indxpath.c discards as irrelevant,
which is a decision made on considerably more information than the
proposed patch uses.

Given the findings above, the two patches are actually complementary. Optimizing the lock fast path not only helps when many indexes exist and only a small subset is used, but whenever there are many locks used by a query. The index filtering is another way to reduce lock contention, but beyond that also greatly reduces the time spent on planning in the serial case.

I have attached the patch to improve the heavyweight lock fast path. It also for now contains moving out _bt_getrootheight(). For workloads where the same set of locks is used over and over again, it only needs on average a single loop iteration to find the relation (instead of a linear scan before). This allows to increase the number of fast path locks by a lot. In this patch I increased them from 16 to 64. The code can be further improved for cases where to be locked relations change frequently and therefore the chance of not finding a relation and because of that having to linearly search the whole array is higher.

I would really appreciate your feedback Tom, also on the questions around the approach of filtering out indexes, discussed in the last mails.

--
David Geier
(ServiceNow)

On 2022-Aug-19, David Geier wrote:

> Beyond that I did some off-CPU profiling to precisely track down which lock
> serializes execution. It turned out to be the MyProc::fpInfoLock lightweight
> lock. This lock is used in the fast path of the heavyweight lock. In the
> contenting case, fpInfoLock is acquired in LW_EXCLUSIVE mode to (1) check if
> there is no other process holding a stronger lock, and if not, to reserve a
> process local fast path lock slot and (2) to return the fast path lock slots
> all in one go. To do so, the current implementation always linearly iterates
> over all lock slots.

Ah, so this is the aspect that you mentioned to me today.  I definitely
think that this analysis deserves its own thread, and the fix is its own
separate patch.

> I have attached the patch to improve the heavyweight lock fast path. It also
> for now contains moving out _bt_getrootheight(). For workloads where the
> same set of locks is used over and over again, it only needs on average a
> single loop iteration to find the relation (instead of a linear scan
> before). This allows to increase the number of fast path locks by a lot. In
> this patch I increased them from 16 to 64. The code can be further improved
> for cases where to be locked relations change frequently and therefore the
> chance of not finding a relation and because of that having to linearly
> search the whole array is higher.

I suggest to put each change in a separate patch:

1. improve fast-path lock algorithm to find the element, perhaps
  together with increasing the number of elements in the array
2. change _bt_getrootheight

However, since patch (1) may have nontrivial performance implications,
you would also need to justify the change: not only that improves the
case where many locks are acquired, but also that it does not make the
case with few locks worse.

I strongly suggest to not include C++ comments or any other dirtiness in
the patch, as that might deter some potential reviewers.

-- 
Álvaro Herrera         PostgreSQL Developer  —  https://www.EnterpriseDB.com/
"It takes less than 2 seconds to get to 78% complete; that's a good sign.
A few seconds later it's at 90%, but it seems to have stuck there.  Did
somebody make percentages logarithmic while I wasn't looking?"
                http://smylers.hates-software.com/2005/09/08/1995c749.html