Memory consumption during partitionwise join planning

Hi All,

When we implemented partitionwise join we disabled it by default (by
setting enable_partitionwise_join to false by default) since it
consumed a lot of memory and took a lot of time [1]. We also set
enable_partitionwise_aggregate to false by default since partitionwise
aggregates require partitionwise join. I have come across at least a
few cases in the field where users didn't know about these GUCs and
suffered from bad performance when queries involved partitioned
tables. Their queries improved a lot by just turning these GUCs ON.
Given that the partitionwise operations improve performance of queries
on partitioned tables, I think these GUCs need to be turned ON by
default.

Irrespective of whether the GUC is turned ON or OFF by default, we
need to reduce the memory consumed by the partitionwise planning and
the time those strategies take during planning. In this email I will
discuss the planner memory consumption problem.

Please find below the memory consumption measurements on master.

Experiment
----------
Created two tables t1 and t1_parted respectively. t1 is an
unpartitioned table whereas t1_parted is a partitioned table with 1000
partitions. Both of them are empty. Being empty does not affect
planning time and memory consumption much, but it takes far less time
to set up and then run EXPLAIN ANALYZE. I am using a self-join on the
partition key to measure memory and planning time. Again it's
something that makes setup easier yet good enough for measurements.
Attached are two scripts. setup.sql should be run to create tables,
partitions and helper functions. queries.sql executes and measures
time and memory of 2-way, 3-way, 4-way and 5-way self joins
respectively. It runs each query four times, once to warm up caches
and then thrice for measurement. Below I report averages of three
runs.

The first attached patch is used to measure memory consumption and
report it as part of EXPLAIN ANALYZE. It simply takes the difference
between memory used in the CurrentMemoryContext before and after
calling pg_plan_query() in ExplainOneQuery(). It does not account for
the memory consumed outside the CurrentMemoryContext while planning
the query e.g. caches. The memory consumed in those contexts does not
necessarily contribute to planning of a given query. Hence they are
ignored. I think the patch by itself makes a good improvement to
EXPLAIN ANALYZE, so we may want to consider it to be accepted in the
core code.

Measurements
------------
In all the tables below, the first column shows the number of tables
joined. Second column is for queries on unpartitioned table. Third
column for queries on partitioned table with partitionwise join turned
OFF. Fourth column is for queries on partitioned table with
partitionwise join turned ON.

Table 1: Planning time in ms.

Number of | unpartitioned | partitioned with | partitioned |
tables | | PWJ off | PWJ on |
---------------------------------------------------------------
2 | 0.10 | 464.39 | 501.75 |
3 | 0.17 | 2,714.28 | 3,035.07 |
4 | 0.34 | 8,856.82 | 10,416.31 |
5 | 0.84 | 22,519.02 | 29,769.19 |

Table 2: Memory consumption

Number of | unpartitioned | partitioned with | partitioned |
tables | | PWJ off | PWJ on |
---------------------------------------------------------------
2 | 29.056 KiB | 19.520 MiB | 40.298 MiB |
3 | 79.088 KiB | 45.227 MiB | 146.877 MiB |
4 | 208.552 KiB | 83.540 MiB | 445.453 MiB |
5 | 561.592 KiB | 149.283 MiB | 1563.253 MiB |

The time required to plan queries on partitioned table is very high
compared to the time taken to plan queries on unpartitioned tables. In
fact it's way higher than the factor of 1000 expected when there 1000
partitions. But the difference between planning time when
partitionwise join is turned OFF vs ON is less compared to difference
between unpartitioned vs partitioned case. The problem of reducing
planning time for queries on partitioned tables is being worked on by
Yuya and David in [2]. That work focuses on queries without
partitionwise join. But it might fix the problem for partitionwise
join as well.

I am focusing on reducing the memory consumption when partitionwise
join is enabled. As visible from table 2 when partitionwise join is
*not* used, the memory consumed in planning is proportional to the
number of partitions as expected. But when partitionwise join is
enabled the memory consumption increases exponentially with the number
of tables being joined. This is because the number of ways a join can
be computed is exponentially proportional to the number of tables
being joined. Table 3 has a breakup of memory consumed.

The memory consumption is broken by the objects that consume memory
during planning. The second attached patch is used to measure breakup
by functionality . Here's a brief explanation of the rows in the
table.

1. Restrictlist translations: Like other expressions the Restrictinfo
lists of parent are translated to obtain Restrictinfo lists to be
applied to child partitions (base as well as join). The first row
shows the memory consumed by the translated RestrictInfos. We can't
avoid these translations but closer examination reveals that a given
RestrictInfo gets translated multiple times proportional to the join
orders. These repeated translations can be avoided. I will start a
separate thread to discuss this topic.

2. Paths: this is the memory consumed when creating child join paths
and the Append paths in parent joins. It includes memory consumed by
the paths as well as translated expressions. I don't think we can
avoid creating these paths. But once the best paths are chosen for the
lower level relations, the unused paths can be freed. I will start a
separate thread to discuss this topic.

3. targetlist translation: child join relations' targetlists are
created by translating parent relations' targetlist. This row shows
the memory consumed by the translated targetlists. This translation
can't be avoided.

4. child SpecialJoinInfo: This is memory consumed in child joins'
SpecialJoinInfos translated from SpecialJoinInfo applicable to parent
joins. The child SpecialJoinInfos are translated on the fly when
computing child joins but are never freed. May be we can free them on
the fly as well or even better save them somewhere and fetch as and
when required. I will start a separate thread to discuss this topic.

5. Child join RelOptInfos: memory consumed by child join relations.
This is unavoidable as we need the RelOptInfos representing the child
joins.

Table 3: Partitionwise join planning memory breakup
Num joins | 2 | 3 | 4 | 5 |
------------------------------------------------------------------------
1. translated | 1.8 MiB | 13.1 MiB | 58.0 MiB | 236.5 MiB |
restrictlists | | | | |
------------------------------------------------------------------------
2. creating child | 11.6 MiB | 59.4 MiB | 207.6 MiB | 768.2 MiB |
join paths | | | | |
------------------------------------------------------------------------
3. translated | 723.5 KiB | 3.3 MiB | 10.6 MiB | 28.5 MiB |
targetlists | | | | |
------------------------------------------------------------------------
4. child | 926.8 KiB | 9.0 MiB | 45.7 MiB | 245.5 MiB |
SpecialJoinInfo | | | | |
------------------------------------------------------------------------
5. Child join rels | 1.6 MiB | 7.9 MiB | 23.8 MiB | 67.5 MiB |
------------------------------------------------------------------------

Rows 1, 2 and 4 show most of the memory consumption. Those are also
the ones where there are opportunities to save memory. As said above,
we will discuss them in their respective email threads since each of
these fixes are independent and require separate discussion.

Here's memory consumption numbers after applying all the POC patches
mentioned above.
Number of tables | no | all POC | % | Absolute |
joined | patches | patches | reduction | reduction |
----------------------------------------------------------------------
2 | 40.3 MiB | 36.5 MiB | 9.43% | 3.8 MiB |
3 | 146.9 MiB | 120.2 MiB | 18.13% | 26.6 MiB |
4 | 445.5 MiB | 328.8 MiB | 26.19% | 116.7 MiB |
5 | 1563.3 MiB | 953.2 MiB | 39.03% | 610.1 MiB |

Even if we are able to recover some memory as mentioned above, a lot
of memory is still consumed by the translated objects (expressions,
RestrictLists etc.). In case of partitioned tables, these translated
object trees differ only in leaf Var nodes. Thus if we device a way to
use the same Var node for parents as well as its children, we won't
need to translate the expressions. But that's a much larger effort and
probably something not even feasible.

While subproblems and their solutions will be discussed in separate
email threads, this thread is to discuss
1. generic memory consumption problems not covered above but seen by others
2. Is the memory consumption reduction after the POC patches good
enough for us to make partitionwise operations turned ON by default?
3. Is there anything else that stops us from turning the partitionwise
operations ON by default?

[1] https://www.postgresql.org/message-id/CA+TgmoYggDp6k-HXNAgrykZh79w6nv2FevpYR_jeMbrORDaQrA@mail.gmail.com
[2] https://www.postgresql.org/message-id/flat/CAJ2pMkZNCgoUKSE%2B_5LthD%2BKbXKvq6h2hQN8Esxpxd%2Bcxmgomg%40mail.gmail.com

--
Best Wishes,
Ashutosh Bapat