The core innovation of this patch is the introduction of the concept
of a partial path; that is, a path which if executed in parallel will
generate a subset of the output rows in each process. Gathering a
partial path produces an ordinary (complete) path. This allows us to
generate paths for parallel joins by joining a partial path for one
side (which at the baserel level is currently always a Partial Seq
Scan) to an ordinary path on the other side. This is subject to
various restrictions at present, especially that this strategy seems
unlikely to be sensible for merge joins, so only nested loops and
hash joins paths are generated.
This also allows an Append node to be pushed below a Gather node in
the case of a partitioned table.
Testing revealed that early versions of this patch made poor decisions
in some cases, which turned out to be caused by the fact that the
original cost model for Parallel Seq Scan wasn't very good. So this
patch tries to make some modest improvements in that area.
There is much more to be done in the area of generating good parallel
plans in all cases, but this seems like a useful step forward.
Patch by me, reviewed by Dilip Kumar and Amit Kapila.
Aggregate nodes now have two new modes: a "partial" mode where they
output the unfinalized transition state, and a "finalize" mode where
they accept unfinalized transition states rather than individual
values as input.
These new modes are not used anywhere yet, but they will be necessary
for parallel aggregation. The infrastructure also figures to be
useful for cases where we want to aggregate local data and remote
data via the FDW interface, and want to bring back partial aggregates
from the remote side that can then be combined with locally generated
partial aggregates to produce the final value. It may also be useful
even when neither FDWs nor parallelism are in play, as explained in
the comments in nodeAgg.c.
David Rowley and Simon Riggs, reviewed by KaiGai Kohei, Heikki
Linnakangas, Haribabu Kommi, and me.
This patch reduces pg_am to just two columns, a name and a handler
function. All the data formerly obtained from pg_am is now provided
in a C struct returned by the handler function. This is similar to
the designs we've adopted for FDWs and tablesample methods. There
are multiple advantages. For one, the index AM's support functions
are now simple C functions, making them faster to call and much less
error-prone, since the C compiler can now check function signatures.
For another, this will make it far more practical to define index access
methods in installable extensions.
A disadvantage is that SQL-level code can no longer see attributes
of index AMs; in particular, some of the crosschecks in the opr_sanity
regression test are no longer possible from SQL. We've addressed that
by adding a facility for the index AM to perform such checks instead.
(Much more could be done in that line, but for now we're content if the
amvalidate functions more or less replace what opr_sanity used to do.)
We might also want to expose some sort of reporting functionality, but
this patch doesn't do that.
Alexander Korotkov, reviewed by Petr Jelínek, and rather heavily
editorialized on by me.
The original coding read tuples from workers in round-robin fashion,
but performance testing shows that it works much better to read enough
to empty one queue before moving on to the next. I believe the
reason for this is that, with the old approach, we could easily wake
up a worker repeatedly to write only one new tuple into the shm_mq
each time. With this approach, by the time the process gets scheduled,
it has a decent chance of being able to fill the entire buffer in
one go.
Patch by me. Dilip Kumar helped with performance testing.
Omitted boundaries represent the upper or lower limit of the corresponding
array subscript. This allows simpler specification of many common
use-cases.
(Revised version of commit 9246af6799)
YUriy Zhuravlev
ExecOnConflictUpdate() passed t_ctid of the to-be-updated tuple to
ExecUpdate(). That's problematic primarily because of two reason: First
and foremost t_ctid could point to a different tuple. Secondly, and
that's what triggered the complaint by Stanislav, t_ctid is changed by
heap_update() to point to the new tuple version. The behavior of AFTER
UPDATE triggers was therefore broken, with NEW.* and OLD.* tuples
spuriously identical within AFTER UPDATE triggers.
To fix both issues, pass a pointer to t_self of a on-stack HeapTuple
instead.
Fixing this bug lead to one change in regression tests, which previously
failed due to the first issue mentioned above. There's a reasonable
expectation that test fails, as it updates one row repeatedly within one
INSERT ... ON CONFLICT statement. That is only possible if the second
update is triggered via ON CONFLICT ... SET, ON CONFLICT ... WHERE, or
by a WITH CHECK expression, as those are executed after
ExecOnConflictUpdate() does a visibility check. That could easily be
prohibited, but given it's allowed for plain UPDATEs and a rare corner
case, it doesn't seem worthwhile.
Reported-By: Stanislav Grozev
Author: Andres Freund and Peter Geoghegan
Discussion: CAA78GVqy1+LisN-8DygekD_Ldfy=BJLarSpjGhytOsgkpMavfQ@mail.gmail.com
Backpatch: 9.5, where ON CONFLICT was introduced
The original parallel sequential scan commit included only very limited
changes to the EXPLAIN output. Aggregated totals from all workers were
displayed, but there was no way to see what each individual worker did
or to distinguish the effort made by the workers from the effort made by
the leader.
Per a gripe by Thom Brown (and maybe others). Patch by me, reviewed
by Amit Kapila.
Commit e7cb7ee145 provided basic
infrastructure for allowing a foreign data wrapper or custom scan
provider to replace a join of one or more tables with a scan.
However, this infrastructure failed to take into account the need
for possible EvalPlanQual rechecks, and ExecScanFetch would fail
an assertion (or just overwrite memory) if such a check was attempted
for a plan containing a pushed-down join. To fix, adjust the EPQ
machinery to skip some processing steps when scanrelid == 0, making
those the responsibility of scan's recheck method, which also has
the responsibility in this case of correctly populating the relevant
slot.
To allow foreign scans to gain control in the right place to make
use of this new facility, add a new, optional RecheckForeignScan
method. Also, allow a foreign scan to have a child plan, which can
be used to correctly populate the slot (or perhaps for something
else, but this is the only use currently envisioned).
KaiGai Kohei, reviewed by Robert Haas, Etsuro Fujita, and Kyotaro
Horiguchi.
The previous coding attempts to destroy the DSM in this case, but
child nodes might have stored data there and still be holding onto
pointers in this case. So don't do that.
Also, free the reader array instead of leaking it.
Extracted from two different patch versions both by Amit Kapila.
When I wrote this code originally, the intention was to recompute the
remapinfo only when the tupledesc changes. This presumably only
happens once per query, but I copied the design pattern from other
DestReceivers. However, due to a silly oversight on my part,
tqueue->tupledesc never got set, leading to recomputation for every
tuple.
This should improve the performance of parallel scans that return a
significant number of tuples.
Report by Amit Kapila; patch by me, reviewed by him.
In addition, this path fills in a number of missing bits and pieces in
the parallel infrastructure. Paths and plans now have a parallel_aware
flag indicating whether whatever parallel-aware logic they have should
be engaged. It is believed that we will need this flag for a number of
path/plan types, not just sequential scans, which is why the flag is
generic rather than part of the SeqScan structures specifically.
Also, execParallel.c now gives parallel nodes a chance to initialize
their PlanState nodes from the DSM during parallel worker startup.
Amit Kapila, with a fair amount of adjustment by me. Review of previous
patch versions by Haribabu Kommi and others.
The patches committed as 6e71dd7ce9
and 3a1f8611f2 were developed in
parallel but dependent on each other in a way that I failed to
notice.
This patch to fix the problem was prepared by Amit Kapila.
Commit 4a4e6893aa, which introduced this
mechanism, failed to account for the fact that the RECORD pseudo-type
uses transient typmods that are only meaningful within a single
backend. Transferring such tuples without modification between two
cooperating backends does not work. This commit installs a system
for passing the tuple descriptors over the same shm_mq being used to
send the tuples themselves. The two sides might not assign the same
transient typmod to any given tuple descriptor, so we must also
substitute the appropriate receiver-side typmod for the one used by
the sender. That adds some CPU overhead, but still seems better than
being unable to pass records between cooperating parallel processes.
Along the way, move the logic for handling multiple tuple queues from
tqueue.c to nodeGather.c; tqueue.c now provides a TupleQueueReader,
which reads from a single queue, rather than a TupleQueueFunnel, which
potentially reads from multiple queues. This change was suggested
previously as a way to make sure that nodeGather.c rather than tqueue.c
had policy control over the order in which to read from queues, but
it wasn't clear to me until now how good an idea it was. typmod
mapping needs to be performed separately for each queue, and it is
much simpler if the tqueue.c code handles that and leaves multiplexing
multiple queues to higher layers of the stack.
Commit d1b7c1ffe7 introduced a mechanism
for serializing a ParamListInfo structure to be passed to a parallel
worker. However, this mechanism failed to handle external expanded
values, as pointed out by Noah Misch. Repair.
Moreover, plpgsql_param_fetch requires adjustment because the
serialization mechanism needs it to skip evaluating unused parameters
just as we would do when it is called from copyParamList, but params
== estate->paramLI in that case. To fix, make the bms_is_member test
in that function unconditional.
Finally, have setup_param_list set a new ParamListInfo field,
paramMask, to the parameters actually used in the expression, so that
we don't try to fetch those that are not needed when serializing a
parameter list. This isn't necessary for correctness, but it makes
the performance of the parallel executor code comparable to what we
do for cases involving cursors.
Design suggestions and extensive review by Noah Misch. Patch by me.
Commit b0b0d84b3d purported to make it
possible to relaunch workers using the same parallel context, but it had
an unpleasant race condition: we might reinitialize after the workers
have sent their last control message but before they have dettached the
DSM, leaving to crashes. Repair by introducing a new ParallelContext
operation, ReinitializeParallelDSM.
Adjust execParallel.c to use this new support, so that we can rescan a
Gather node by relaunching workers but without needing to recreate the
DSM.
Amit Kapila, with some adjustments by me. Extracted from latest parallel
sequential scan patch.
The original Gather code failed to mark a Gather node as not able to
do projection, but it couldn't, even though it did call initialize its
projection info via ExecAssignProjectionInfo. There doesn't seem to
be any good reason for this node not to have projection capability,
so clean things up so that it does. Without this, plans using Gather
nodes might need to carry extra Result nodes to do projection.
Commit 816e336f12 added the wrong error
check to async.c; sending restrictions is restricted to the leader,
not altogether unsafe.
Commit 3bd909b220 added ExecShutdownNode
to traverse the planstate tree and call shutdown functions, but made
a Gather node, the only node that actually has such a function, abort
the tree traversal, which is wrong.
In the previous coding, before returning from ExecutorRun, we'd shut
down all parallel workers. This was dead wrong if ExecutorRun was
called with a non-zero tuple count; it had the effect of truncating
the query output. To fix, give ExecutePlan control over whether to
enter parallel mode, and have it refuse to do so if the tuple count
is non-zero. Rewrite the Gather logic so that it can cope with being
called outside parallel mode.
Commit 7aea8e4f2d is largely to blame
for this problem, though this patch modifies some subsequently-committed
code which relied on the guarantees it purported to make.
This fixes a long-standing bug which was discovered while investigating
the interaction between the new join pushdown code and the EvalPlanQual
machinery: if a ForeignScan appears on the inner side of a paramaterized
nestloop, an EPQ recheck would re-return the original tuple even if
it no longer satisfied the pushed-down quals due to changed parameter
values.
This fix adds a new member to ForeignScan and ForeignScanState and a
new argument to make_foreignscan, and requires changes to FDWs which
push down quals to populate that new argument with a list of quals they
have chosen to push down. Therefore, I'm only back-patching to 9.5,
even though the bug is not new in 9.5.
Etsuro Fujita, reviewed by me and by Kyotaro Horiguchi.
On reflection, the submitted patch didn't really work to prevent the
request size from exceeding MaxAllocSize, because of the fact that we'd
happily round nbuckets up to the next power of 2 after we'd limited it to
max_pointers. The simplest way to enforce the limit correctly is to
round max_pointers down to a power of 2 when it isn't one already.
(Note that the constraint to INT_MAX / 2, if it were doing anything useful
at all, is properly applied after that.)
Limit the size of the hashtable pointer array to not more than
MaxAllocSize, per reports from Kouhei Kaigai and others of "invalid memory
alloc request size" failures. There was discussion of allowing the array
to get larger than that by using the "huge" palloc API, but so far no proof
that that is actually a good idea, and at this point in the 9.5 cycle major
changes from old behavior don't seem like the way to go.
Fix a rather serious secondary bug in the new code, which was that it
didn't ensure nbuckets remained a power of 2 when recomputing it for the
multiple-batch case.
Clean up sloppy division of labor between ExecHashIncreaseNumBuckets and
its sole call site.
A Gather executor node runs any number of copies of a plan in an equal
number of workers and merges all of the results into a single tuple
stream. It can also run the plan itself, if the workers are
unavailable or haven't started up yet. It is intended to work with
the Partial Seq Scan node which will be added in future commits.
It could also be used to implement parallel query of a different sort
by itself, without help from Partial Seq Scan, if the single_copy mode
is used. In that mode, a worker executes the plan, and the parallel
leader does not, merely collecting the worker's results. So, a Gather
node could be inserted into a plan to split the execution of that plan
across two processes. Nested Gather nodes aren't currently supported,
but we might want to add support for that in the future.
There's nothing in the planner to actually generate Gather nodes yet,
so it's not quite time to break out the champagne. But we're getting
close.
Amit Kapila. Some designs suggestions were provided by me, and I also
reviewed the patch. Single-copy mode, documentation, and other minor
changes also by me.
This code provides infrastructure for a parallel leader to start up
parallel workers to execute subtrees of the plan tree being executed
in the master. User-supplied parameters from ParamListInfo are passed
down, but PARAM_EXEC parameters are not. Various other constructs,
such as initplans, subplans, and CTEs, are also not currently shared.
Nevertheless, there's enough here to support a basic implementation of
parallel query, and we can lift some of the current restrictions as
needed.
Amit Kapila and Robert Haas
When taking the UPDATE path in an INSERT .. ON CONFLICT .. UPDATE tables
with oids were not supported. The tuple generated by the update target
list was projected without space for an oid - a simple oversight.
Reported-By: Peter Geoghegan
Author: Andres Freund
Backpatch: 9.5, where ON CONFLICT was introduced
The shm_mq mechanism was built to send error (and notice) messages and
tuples between backends. However, shm_mq itself only deals in raw
bytes. Since commit 2bd9e412f9, we have
had infrastructure for one message to redirect protocol messages to a
queue and for another backend to parse them and do useful things with
them. This commit introduces a somewhat analogous facility for tuples
by adding a new type of DestReceiver, DestTupleQueue, which writes
each tuple generated by a query into a shm_mq, and a new
TupleQueueFunnel facility which reads raw tuples out of the queue and
reconstructs the HeapTuple format expected by the executor.
The TupleQueueFunnel abstraction supports reading from multiple tuple
streams at the same time, but only in round-robin fashion. Someone
could imaginably want other policies, but this should be good enough
to meet our short-term needs related to parallel query, and we can
always extend it later.
This also makes one minor addition to the shm_mq API that didn'
seem worth breaking out as a separate patch.
Extracted from Amit Kapila's parallel sequential scan patch. This
code was originally written by me, and then it was revised by Amit,
and then it was revised some more by me.
Commit 924bcf4f16 introduced a framework
for parallel computation in PostgreSQL that makes most but not all
built-in functions safe to execute in parallel mode. In order to have
parallel query, we'll need to be able to determine whether that query
contains functions (either built-in or user-defined) that cannot be
safely executed in parallel mode. This requires those functions to be
labeled, so this patch introduces an infrastructure for that. Some
functions currently labeled as safe may need to be revised depending on
how pending issues related to heavyweight locking under paralllelism
are resolved.
Parallel plans can't be used except for the case where the query will
run to completion. If portal execution were suspended, the parallel
mode restrictions would need to remain in effect during that time, but
that might make other queries fail. Therefore, this patch introduces
a framework that enables consideration of parallel plans only when it
is known that the plan will be run to completion. This probably needs
some refinement; for example, at bind time, we do not know whether a
query run via the extended protocol will be execution to completion or
run with a limited fetch count. Having the client indicate its
intentions at bind time would constitute a wire protocol break. Some
contexts in which parallel mode would be safe are not adjusted by this
patch; the default is not to try parallel plans except from call sites
that have been updated to say that such plans are OK.
This commit doesn't introduce any parallel paths or plans; it just
provides a way to determine whether they could potentially be used.
I'm committing it on the theory that the remaining parallel sequential
scan patches will also get committed to this release, hopefully in the
not-too-distant future.
Robert Haas and Amit Kapila. Reviewed (in earlier versions) by Noah
Misch.
This refactors rewrite/rowsecurity.c to simplify the handling of the
default deny case (reducing the number of places where we check for and
add the default deny policy from three to one) by splitting up the
retrival of the policies from the application of them.
This also allowed us to do away with the policy_id field. A policy_name
field was added for WithCheckOption policies and is used in error
reporting, when available.
Patch by Dean Rasheed, with various mostly cosmetic changes by me.
Back-patch to 9.5 where RLS was introduced to avoid unnecessary
differences, since we're still in alpha, per discussion with Robert.
Per discussion, nowadays it is possible to have tablespaces that have
wildly different I/O characteristics from others. Setting different
effective_io_concurrency parameters for those has been measured to
improve performance.
Author: Julien Rouhaud
Reviewed by: Andres Freund
For no obvious reason, spi_printtup() was coded to enlarge the tuple
pointer table by just 256 slots at a time, rather than doubling the size at
each reallocation, as is our usual habit. For very large SPI results, this
makes for O(N^2) time spent in repalloc(), which of course soon comes to
dominate the runtime. Use the standard doubling approach instead.
This is a longstanding performance bug, so back-patch to all active
branches.
Neil Conway
Commit 8cce08f168 used a left-shift
on a literal of 1 that could (in large allocations) be shifted by
31 or more bits. This was assigned to a local variable that was
already declared to be a long to protect against overruns of int,
but the literal in this shift needs to be declared long to allow it
to work correctly in some compilers.
Backpatch to 9.5, where the bug was introduced.
Report and patch by KaiGai Kohei, slighly modified based on
discussion.
Robert Haas