All the tape functions, like LogicalTapeRead and LogicalTapeWrite, now
take a LogicalTape as argument, instead of LogicalTapeSet+tape number.
You can create any number of LogicalTapes in a single LogicalTapeSet, and
you don't need to decide the number upfront, when you create the tape set.
This makes the tape management in hash agg spilling in nodeAgg.c simpler.
Discussion: https://www.postgresql.org/message-id/420a0ec7-602c-d406-1e75-1ef7ddc58d83%40iki.fi
Reviewed-by: Peter Geoghegan, Zhihong Yu, John Naylor
The point of introducing the hash_mem_multiplier GUC was to let users
reproduce the old behavior of hash aggregation, i.e. that it could use
more than work_mem at need. However, the implementation failed to get
the job done on Win64, where work_mem is clamped to 2GB to protect
various places that calculate memory sizes using "long int". As
written, the same clamp was applied to hash_mem. This resulted in
severe performance regressions for queries requiring a bit more than
2GB for hash aggregation, as they now spill to disk and there's no
way to stop that.
Getting rid of the work_mem restriction seems like a good idea, but
it's a big job and could not conceivably be back-patched. However,
there's only a fairly small number of places that are concerned with
the hash_mem value, and it turns out to be possible to remove the
restriction there without too much code churn or any ABI breaks.
So, let's do that for now to fix the regression, and leave the
larger task for another day.
This patch does introduce a bit more infrastructure that should help
with the larger task, namely pg_bitutils.h support for working with
size_t values.
Per gripe from Laurent Hasson. Back-patch to v13 where the
behavior change came in.
Discussion: https://postgr.es/m/997817.1627074924@sss.pgh.pa.us
Discussion: https://postgr.es/m/MN2PR15MB25601E80A9B6D1BA6F592B1985E39@MN2PR15MB2560.namprd15.prod.outlook.com
Here we alter the code that calls build_pertrans_for_aggref() so that the
function no longer needs to special-case whether it's dealing with an
aggtransfn or an aggcombinefn. This allows us to reuse the
build_aggregate_transfn_expr() function and just get rid of the
build_aggregate_combinefn_expr() completely.
All of the special case code that was in build_pertrans_for_aggref() has
been moved up to the calling functions.
This saves about a dozen lines of code in nodeAgg.c and a few dozen more
in parse_agg.c
Also, rename a few variables in nodeAgg.c to try to make it more clear
that we're working with either a aggtransfn or an aggcombinefn. Some of
the old names would have you believe that we were always working with an
aggtransfn.
Discussion: https://postgr.es/m/CAApHDvptMQ9FmF0D67zC_w88yVnoNVR2+kkOQGUrCmdxWxLULQ@mail.gmail.com
Also "make reformat-dat-files".
The only change worthy of note is that pgindent messed up the formatting
of launcher.c's struct LogicalRepWorkerId, which led me to notice that
that struct wasn't used at all anymore, so I just took it out.
Commit 230230223 taught nodeAgg.c that, when spilling tuples from
memory in an oversized hash aggregation, it only needed to spill
input columns referenced in the node's tlist and quals. Unfortunately,
that's wrong: we also have to save the grouping columns. The error
is masked in common cases because the grouping columns also appear
in the tlist, but that's not necessarily true. The main category
of plans where it's not true seem to come from semijoins ("WHERE
outercol IN (SELECT innercol FROM innertable)") where the innercol
needs an implicit promotion to make it comparable to the outercol.
The grouping column will be "innercol::promotedtype", but that
expression appears nowhere in the Agg node's own tlist and quals;
only the bare "innercol" is found in the tlist.
I spent quite a bit of time looking for a suitable regression test
case for this, without much success. If the number of distinct
values of the innercol is large enough to make spilling happen,
the planner tends to prefer a non-HashAgg plan, at least for
problem sizes that are reasonable to use in the regression tests.
So, no new regression test. However, this patch does demonstrably
fix the originally-reported test case.
Per report from s.p.e (at) gmx-topmail.de. Backpatch to v13
where the troublesome code came in.
Discussion: https://postgr.es/m/trinity-1c565d44-159f-488b-a518-caf13883134f-1611835701633@3c-app-gmx-bap78
Before processing tuples, agg_refill_hash_table() was setting all
pergroup pointers to NULL to signal to advance_aggregates() that it
should not attempt to advance groups that had spilled.
The problem was that it also set the pergroups for sorted grouping
sets to NULL, which caused rescanning to fail.
Instead, change agg_refill_hash_table() to only set the pergroups for
hashed grouping sets to NULL; and when compiling the expression, pass
doSort=false.
Reported-by: Alexander Lakhin
Discussion: https://postgr.es/m/16784-7ff169bf2c3d1588%40postgresql.org
Backpatch-through: 13
Gen_fmgrtab.pl treated aggregate functions the same as other built-in
functions, which is wasteful because there is no real need to have
entries for them in the fmgr_builtins[] table. Suppressing those
entries saves about 3KB in the compiled table on my machine; which
is not a lot but it's not nothing either, considering that that
table is pretty "hot". The only outside code change needed is
that ExecInitWindowAgg() can't be allowed to call fmgr_info_cxt()
on a plain aggregate function. But that saves a few cycles anyway.
Having done that, the aggregate_dummy() function is unreferenced
and might as well be dropped. Using "aggregate_dummy" as the prosrc
value for an aggregate is now just a documentation convention not
something that matters. There was some discussion of using NULL
instead to save a few bytes in pg_proc, but we'd have to remove
prosrc's BKI_FORCE_NOT_NULL marking which doesn't seem a great idea.
Anyway, it's possible there's client-side code that expects to
see "aggregate_dummy" there, so I'm loath to change it without a
strong reason.
Discussion: https://postgr.es/m/533989.1604263665@sss.pgh.pa.us
Previously, it was based on nBlocksAllocated to account for tapes with
open write buffers that may not have made it to the BufFile yet.
That was unnecessary, because callers do not need to get the number of
blocks while a tape has an open write buffer; and it also conflicted
with the preallocation logic added for HashAgg.
Reviewed-by: Peter Geoghegan
Discussion: https://postgr.es/m/ce5af05900fdbd0e9185747825a7423c48501964.camel@j-davis.com
Backpatch-through: 13
Add a GUC that acts as a multiplier on work_mem. It gets applied when
sizing executor node hash tables that were previously size constrained
using work_mem alone.
The new GUC can be used to preferentially give hash-based nodes more
memory than the generic work_mem limit. It is intended to enable admin
tuning of the executor's memory usage. Overall system throughput and
system responsiveness can be improved by giving hash-based executor
nodes more memory (especially over sort-based alternatives, which are
often much less sensitive to being memory constrained).
The default value for hash_mem_multiplier is 1.0, which is also the
minimum valid value. This means that hash-based nodes continue to apply
work_mem in the traditional way by default.
hash_mem_multiplier is generally useful. However, it is being added now
due to concerns about hash aggregate performance stability for users
that upgrade to Postgres 13 (which added disk-based hash aggregation in
commit 1f39bce0). While the old hash aggregate behavior risked
out-of-memory errors, it is nevertheless likely that many users actually
benefited. Hash agg's previous indifference to work_mem during query
execution was not just faster; it also accidentally made aggregation
resilient to grouping estimate problems (at least in cases where this
didn't create destabilizing memory pressure).
hash_mem_multiplier can provide a certain kind of continuity with the
behavior of Postgres 12 hash aggregates in cases where the planner
incorrectly estimates that all groups (plus related allocations) will
fit in work_mem/hash_mem. This seems necessary because hash-based
aggregation is usually much slower when only a small fraction of all
groups can fit. Even when it isn't possible to totally avoid hash
aggregates that spill, giving hash aggregation more memory will reliably
improve performance (the same cannot be said for external sort
operations, which appear to be almost unaffected by memory availability
provided it's at least possible to get a single merge pass).
The PostgreSQL 13 release notes should advise users that increasing
hash_mem_multiplier can help with performance regressions associated
with hash aggregation. That can be taken care of by a later commit.
Author: Peter Geoghegan
Reviewed-By: Álvaro Herrera, Jeff Davis
Discussion: https://postgr.es/m/20200625203629.7m6yvut7eqblgmfo@alap3.anarazel.de
Discussion: https://postgr.es/m/CAH2-WzmD%2Bi1pG6rc1%2BCjc4V6EaFJ_qSuKCCHVnH%3DoruqD-zqow%40mail.gmail.com
Backpatch: 13-, where disk-based hash aggregation was introduced.
Use HyperLogLog to estimate the group cardinality in a spilled
partition. This estimate is used to choose the number of partitions if
we recurse.
The previous behavior was to use the number of tuples in a spilled
partition as the estimate for the number of groups, which lead to
overpartitioning. That could cause the number of batches to be much
higher than expected (with each batch being very small), which made it
harder to interpret EXPLAIN ANALYZE results.
Reviewed-by: Peter Geoghegan
Discussion: https://postgr.es/m/a856635f9284bc36f7a77d02f47bbb6aaf7b59b3.camel@j-davis.com
Backpatch-through: 13
There were various unnecessary differences between Hash Agg's EXPLAIN
ANALYZE output and Hash Join's. Here we modify the Hash Agg output so
that it's better aligned to Hash Join's.
The following changes have been made:
1. Start batches counter at 1 instead of 0.
2. Always display the "Batches" property, even when we didn't spill to
disk.
3. Use the text "Batches" instead of "HashAgg Batches" for text format.
4. Use the text "Memory Usage" instead of "Peak Memory Usage" for text
format.
5. Include "Batches" before "Memory Usage" in both text and non-text
formats.
In passing also modify the "Planned Partitions" property so that we show
it regardless of if the value is 0 or not for non-text EXPLAIN formats.
This was pointed out by Justin Pryzby and probably should have been part
of 40efbf870.
Reviewed-by: Justin Pryzby, Jeff Davis
Discussion: https://postgr.es/m/CAApHDvrshRnA6C0VFnu7Fb9TVvgGo80PUMm5+2DiaS1gEkPvtw@mail.gmail.com
Backpatch-through: 13, where HashAgg batching was introduced
The term "hash_mem" will take on new significance when pending work to
add a new hash_mem_multiplier GUC is committed. Rename a local variable
that happens to have been called hash_mem now to avoid confusion.
Since 1f39bce02, HashAgg nodes have had the ability to spill to disk when
memory consumption exceeds work_mem. That commit added new properties to
EXPLAIN ANALYZE to show the maximum memory usage and disk usage, however,
it didn't quite go as far as showing that information for parallel
workers. Since workers may have experienced something very different from
the main process, we should show this information per worker, as is done
in Sort.
Reviewed-by: Justin Pryzby
Reviewed-by: Jeff Davis
Discussion: https://postgr.es/m/CAApHDvpEKbfZa18mM1TD7qV6PG+w97pwCWq5tVD0dX7e11gRJw@mail.gmail.com
Backpatch-through: 13, where the hashagg spilling code was added.
Thomas Munro fixed a longstanding annoyance in pg_bsd_indent, that
it would misformat lines containing IsA() macros on the assumption
that the IsA() call should be treated like a cast. This improves
some other cases involving field/variable names that match typedefs,
too. The only places that get worse are a couple of uses of the
OpenSSL macro STACK_OF(); we'll gladly take that trade-off.
Discussion: https://postgr.es/m/20200114221814.GA19630@alvherre.pgsql
Includes some manual cleanup of places that pgindent messed up,
most of which weren't per project style anyway.
Notably, it seems some people didn't absorb the style rules of
commit c9d297751, because there were a bunch of new occurrences
of function calls with a newline just after the left paren, all
with faulty expectations about how the rest of the call would get
indented.
Working on commit 1c455078b led me to check through FunctionCallInvoke
call sites to see if every one was being honest about (a) making sure
that fcinfo.isnull is initially false, and (b) checking its state after
the call. Sure enough, I found some violations.
The main one is that finalize_partialaggregate re-used serialfn_fcinfo
without resetting isnull, even though it clearly intends to cater for
serialfns that return NULL. There would only be an issue with a
non-strict serialfn, since it's unlikely that a serialfn would return
NULL for non-null input. We have no non-strict serialfns in core, and
there may be none in the wild either, which would account for the lack
of complaints. Still, it's clearly wrong, so back-patch that fix to
9.6 where finalize_partialaggregate was introduced.
Also, arrayfuncs.c and rowtypes.c contained various callers that were
not bothering to check for result nulls. While what's being called is
a comparison or hash function that probably *shouldn't* return null,
that's a lousy excuse for not having any check at all. There are
existing places that just Assert(!fcinfo->isnull) in comparable
situations, so I added that to the places that were calling btree
comparison or hash support functions. In the places calling
boolean-returning equality functions, it's quite cheap to have them
treat isnull as FALSE, so make those places do that. Also remove some
"locfcinfo->isnull = false" assignments that are unnecessary given the
assumption that no previous call returned null. These changes seem like
mostly neatnik-ism or debugging support, so I didn't back-patch.
If the memory context's maxBlockSize is too big, a single block
allocation can suddenly exceed work_mem. For Hash Aggregation, this
can mean spilling to disk too early or reporting a confusing memory
usage number for EXPLAN ANALYZE.
Introduce CreateWorkExprContext(), which is like CreateExprContext(),
except that it creates the AllocSet with a maxBlockSize that is
reasonable in proportion to work_mem.
Right now, CreateWorkExprContext() is only used by Hash Aggregation,
but it may be generally useful in the future.
Discussion: https://postgr.es/m/412a3fbf306f84d8d78c4009e11791867e62b87c.camel@j-davis.com
Don't try to be precise about it, just use a constant 16 bytes of
chunk overhead. Being smarter would require knowing the memory context
where the chunk will be allocated, which is not known by all callers.
Discussion: https://postgr.es/m/20200325220936.il3ni2fj2j2b45y5@alap3.anarazel.de
While performing hash aggregation, track memory usage when adding new
groups to a hash table. If the memory usage exceeds work_mem, enter
"spill mode".
In spill mode, new groups are not created in the hash table(s), but
existing groups continue to be advanced if input tuples match. Tuples
that would cause a new group to be created are instead spilled to a
logical tape to be processed later.
The tuples are spilled in a partitioned fashion. When all tuples from
the outer plan are processed (either by advancing the group or
spilling the tuple), finalize and emit the groups from the hash
table. Then, create new batches of work from the spilled partitions,
and select one of the saved batches and process it (possibly spilling
recursively).
Author: Jeff Davis
Reviewed-by: Tomas Vondra, Adam Lee, Justin Pryzby, Taylor Vesely, Melanie Plageman
Discussion: https://postgr.es/m/507ac540ec7c20136364b5272acbcd4574aa76ef.camel@j-davis.com
Optionally push a step to check for a NULL pointer to the pergroup
state.
This will be important for disk-based hash aggregation in combination
with grouping sets. When memory limits are reached, a given tuple may
find its per-group state for some grouping sets but not others. For
the former, it advances the per-group state as normal; for the latter,
it skips evaluation and the calling code will have to spill the tuple
and reprocess it in a later batch.
Add the NULL check as a separate expression step because in some
common cases it's not needed.
Discussion: https://postgr.es/m/20200221202212.ssb2qpmdgrnx52sj%40alap3.anarazel.de
Do so by combining the various steps that are part of aggregate
transition function invocation into one larger step. As some of the
current steps are only necessary for some aggregates, have one variant
of the aggregate transition step for each possible combination.
To avoid further manual copies of code in the different transition
step implementations, move most of the code into helper functions
marked as "always inline".
The benefit of this change is an increase in performance when
aggregating lots of rows. This comes in part due to the reduced number
of indirect jumps due to the reduced number of steps, and in part by
reducing redundant setup code across steps. This mainly benefits
interpreted execution, but the code generated by JIT is also improved
a bit.
As a nice side-effect it also ends up making the code a bit simpler.
A small additional optimization is removing the need to set
aggstate->curaggcontext before calling ExecAggInitGroup, choosing to
instead passign curaggcontext as an argument. It was, in contrast to
other aggregate related functions, only needed to fetch a memory
context to copy the transition value into.
Author: Andres Freund
Discussion:
https://postgr.es/m/20191023163849.sosqbfs5yenocez3@alap3.anarazel.dehttps://postgr.es/m/5c371df7cee903e8cd4c685f90c6c72086d3a2dc.camel@j-davis.com
* Separate calculation of hash value from the lookup.
* Split build_hash_table() into two functions.
* Change lookup_hash_entry() to return AggStatePerGroup. That's all
the caller needed, anyway.
These changes are to support the upcoming Disk-based Hash Aggregation
work.
Discussion: https://postgr.es/m/31f5ab871a3ad5a1a91a7a797651f20e77ac7ce3.camel%40j-davis.com
We used to strategically place newlines after some function call left
parentheses to make pgindent move the argument list a few chars to the
left, so that the whole line would fit under 80 chars. However,
pgindent no longer does that, so the newlines just made the code
vertically longer for no reason. Remove those newlines, and reflow some
of those lines for some extra naturality.
Reviewed-by: Michael Paquier, Tom Lane
Discussion: https://postgr.es/m/20200129200401.GA6303@alvherre.pgsql
The code checking whether an aggregate transition value needs to be
reparented into the current context has always only compared the
transition return value with the previous transition value by datum,
i.e. without regard for NULLness. This normally works, because when
the transition function returns NULL (via fcinfo->isnull), it'll
return a value that won't be the same as its input value.
But there's no hard requirement that that's the case. And it turns
out, it's possible to hit this case (see discussion or reproducers),
leading to a non-null transition value not being reparented, followed
by a crash caused by that.
Instead of adding another comparison of NULLness, instead have
ExecAggTransReparent() ensure that pergroup->transValue ends up as 0
when the new transition value is NULL. That avoids having to add an
additional branch to the much more common cases of the transition
function returning the old transition value (which is a pointer in
this case), and when the new value is different, but not NULL.
In branches since 69c3936a14, also deduplicate the reparenting code
between the expression evaluation based transitions, and the path for
ordered aggregates.
Reported-By: Teodor Sigaev, Nikita Glukhov
Author: Andres Freund
Discussion: https://postgr.es/m/bd34e930-cfec-ea9b-3827-a8bc50891393@sigaev.ru
Backpatch: 9.4-, this issue has existed since at least 7.4
For most uses of acl.h the details of how "Acl" internally looks like
are irrelevant. It might make sense to move a lot of the
implementation details into a separate header at a later point.
The main motivation of this change is to avoid including fmgr.h (via
array.h, which needs it for exposed structs) in a lot of files that
otherwise don't need it. A subsequent commit will remove the fmgr.h
include from a lot of files.
Directly include utils/array.h and utils/expandeddatum.h from the
files that need them, but previously included them indirectly, via
acl.h.
Author: Andres Freund
Discussion: https://postgr.es/m/20190803193733.g3l3x3o42uv4qj7l@alap3.anarazel.de
Since 15d8f8312 we assert that - and since 7ef04e4d2c, 4da597edf1
rely on - the slot type for an expression's
ecxt_{outer,inner,scan}tuple not changing, unless explicitly flagged
as such. That allows to either skip deforming (for a virtual tuple
slot) or optimize the code for JIT accelerated deforming
appropriately (for other known slot types).
This assumption was sometimes violated for grouping sets, when
nodeAgg.c internally uses tuplesorts, and the child node doesn't
return a TTSOpsMinimalTuple type slot. Detect that case, and flag that
the outer slot might not be "fixed".
It's probably worthwhile to optimize this further in the future, and
more granularly determine whether the slot is fixed. As we already
instantiate per-phase transition and equal expressions, we could
cheaply set the slot type appropriately for each phase. But that's a
separate change from this bugfix.
This commit does include a very minor optimization by avoiding to
create a slot for handling tuplesorts, if no such sorts are
performed. Previously we created that slot unnecessarily in the common
case of computing all grouping sets via hashing. The code looked too
confusing without that, as the conditions for needing a sort slot and
flagging that the slot type isn't fixed, are the same.
Reported-By: Ashutosh Sharma
Author: Andres Freund
Discussion: https://postgr.es/m/CAE9k0PmNaMD2oHTEAhRyxnxpaDaYkuBYkLa1dpOpn=RS0iS2AQ@mail.gmail.com
Backpatch: 12-, where the bug was introduced in 15d8f8312
This is numbered take 7, and addresses a set of issues around:
- Fixes for typos and incorrect reference names.
- Removal of unneeded comments.
- Removal of unreferenced functions and structures.
- Fixes regarding variable name consistency.
Author: Alexander Lakhin
Discussion: https://postgr.es/m/10bfd4ac-3e7c-40ab-2b2e-355ed15495e8@gmail.com
When there were duplicate columns in the hash key list, the array
sizes could be miscomputed, resulting in access off the end of the
array. Adjust the computation to ensure the array is always large
enough.
(I considered whether the duplicates could be removed in planning, but
I can't rule out the possibility that duplicate columns might have
different hash functions assigned. Simpler to just make sure it works
at execution time regardless.)
Bug apparently introduced in fc4b3dea2 as part of narrowing down the
tuples stored in the hashtable. Reported by Colm McHugh of Salesforce,
though I didn't use their patch. Backpatch back to version 10 where
the bug was introduced.
Discussion: https://postgr.es/m/CAFeeJoKKu0u+A_A9R9316djW-YW3-+Gtgvy3ju655qRHR3jtdA@mail.gmail.com
This is still using the 2.0 version of pg_bsd_indent.
I thought it would be good to commit this separately,
so as to document the differences between 2.0 and 2.1 behavior.
Discussion: https://postgr.es/m/16296.1558103386@sss.pgh.pa.us
For partial aggregation combine steps,
AggStatePerTrans->numTransInputs was set to the transition function's
number of inputs, rather than the combine function's number of
inputs (always 1).
That lead to partial aggregates with strict combine functions to
wrongly check for NOT NULL input as required by strictness. When the
aggregate wasn't exactly passed one argument, the strictness check was
either omitted (in the 0 args case) or too many arguments were
checked. In the latter case we'd read beyond the end of
FunctionCallInfoData->args (only in master).
AggStatePerTrans->numTransInputs actually has been wrong since since
9.6, where partial aggregates were added. But it turns out to not be
an active problem in 9.6 and 10, because numTransInputs wasn't used at
all for combine functions: Before c253b722f6 there simply was no NULL
check for the input to strict trans functions, and after that the
check was simply hardcoded for the right offset in fcinfo, as it's
done by code specific to combine functions.
In bf6c614a2f (11) the strictness check was generalized, with common
code doing the strictness checks for both plain and combine transition
functions, based on numTransInputs. For combine functions this lead to
not emitting an expression step to check for strict input in the 0
arguments case, and in the > 1 arguments case, we'd check too many
arguments.Due to the fact that the relevant fcinfo->isnull[2..] was
always zero-initialized (more or less by accident, by being part of
the AggStatePerTrans struct, which is palloc0'ed), there was no
observable damage in the latter case before a9c35cf85c, we just
checked too many array elements.
Due to the changes in a9c35cf85c, > 1 argument bug became visible,
because these days fcinfo is a) dynamically allocated without being
zeroed b) exactly the length required for the number of specified
arguments (hardcoded to 2 in this case).
This commit only contains a fairly minimal fix, setting numTransInputs
to a hardcoded 1 when building a pertrans for a combine function. It
seems likely that we'll want to clean this up further (e.g. the
arguments build_pertrans_for_aggref() aren't particularly meaningful
for combine functions). But the wrap date for 12 beta1 is coming up
fast, so it seems good to have a minimal fix in place.
Backpatch to 11. While AggStatePerTrans->numTransInputs was set
wrongly before that, the value was not used for combine functions.
Reported-By: Rajkumar Raghuwanshi
Diagnosed-By: Kyotaro Horiguchi, Jeevan Chalke, Andres Freund, David Rowley
Author: David Rowley, Kyotaro Horiguchi, Andres Freund
Discussion: https://postgr.es/m/CAKcux6=uZEyWyLw0N7HtR9OBc-sWEFeByEZC7t-KDf15FKxVew@mail.gmail.com
Heikki Linnakangas
Tom Lane
David Rowley
Michael Paquier
Bruce Momjian
Jeff Davis
Peter Geoghegan
Andres Freund
Alvaro Herrera
Amit Kapila
Andrew Gierth