The optimization for inserts into BRIN indexes added by c1ec02be1d
relies on a cache that needs to be explicitly released after calling
index_insert(). The commit however failed to invoke the cleanup in
validate_index(), which calls index_insert() indirectly through
table_index_validate_scan().
After inspecting index_insert() callers, it seems unique_key_recheck()
is missing the call too.
Fixed by adding the two missing index_insert_cleanup() calls.
The commit does two additional improvements. The aminsertcleanup()
signature is modified to have the index as the first argument, to make
it more like the other AM callbacks. And the aminsertcleanup() callback
is invoked even if the ii_AmCache is NULL, so that it can decide if the
cleanup is necessary.
Author: Alvaro Herrera, Tomas Vondra
Reported-by: Alexander Lakhin
Discussion: https://postgr.es/m/202401091043.e3nrqiad6gb7@alvherre.pgsql
This fixes various typos, duplicated words, and tiny bits of whitespace
mainly in code comments but also in docs.
Author: Daniel Gustafsson <daniel@yesql.se>
Author: Heikki Linnakangas <hlinnaka@iki.fi>
Author: Alexander Lakhin <exclusion@gmail.com>
Author: David Rowley <dgrowleyml@gmail.com>
Author: Nazir Bilal Yavuz <byavuz81@gmail.com>
Discussion: https://postgr.es/m/3F577953-A29E-4722-98AD-2DA9EFF2CBB8@yesql.se
Preprocessing for nbtree index scans allowed array "input" scan keys
already marked eliminated during array-specific preprocessing to be
"fixed up" during preprocessing proper. This allowed eliminated scan
keys on DESC index columns to spurious have their strategy commuted,
causing assertion failures.
To fix, teach _bt_fix_scankey_strategy to ignore these scan keys. This
brings it in line with its only caller, _bt_preprocess_keys.
Oversight in commit 5bf748b8, which enhanced nbtree ScalarArrayOp
execution.
Reported-By: Donghang Lin <donghanglin@gmail.com>
Discussion: https://postgr.es/m/CAA=D8a2sHK6CAzZ=0CeafC-Y-MFXbYxnRSHvZTi=+JHu6kAa8Q@mail.gmail.com
fefd9a3fed turned tail recursion of CommitTransactionCommand() and
AbortCurrentTransaction() into iteration. However, it splits the handling of
cases between different functions.
This commit puts the handling of all the cases into
AbortCurrentTransactionInternal() and CommitTransactionCommandInternal().
Now CommitTransactionCommand() and AbortCurrentTransaction() are just doing
the repeated calls of internal functions.
Reported-by: Andres Freund
Discussion: https://postgr.es/m/20240415224834.w6piwtefskoh32mv%40awork3.anarazel.de
Author: Andres Freund
Commit b437571714 added support for parallel builds of BRIN indexes,
using code similar to BTREE. But there were to be a couple unnecessary
differences, particularly in how the leader waits for the workers, and
merges the results. So remove these, to make the code more similar.
The leader never waited on the workersdonecv condition variable, but
simply called WaitForParallelWorkersToFinish() in _brin_end_parallel()
and then merged the per-worker results. This worked correctly, but it
seems better to do the wait and merge before _brin_end_parallel().
This commit moves the relevant code to _brin_parallel_heapscan/merge(),
which means _brin_end_parallel() remains responsible only for exiting
the parallel mode and accumulating WAL usage data.
Discussion: https://postgr.es/m/3733d042-71e1-6ae6-5fac-00c12db62db6@enterprisedb.com
Commit 6bcda4a721 replaced PG_DETOAST_DATUM with PG_DETOAST_DATUM_PACKED
in two BRIN output functions, for minmax-multi and bloom opclasses. But
this is incorrect - the code is accessing the data through structs that
already include a 4B header, so the detoast needs to match that. But the
PACKED macro may keep the 1B header, which means the struct fields will
point to incorrect data.
Backpatch-through: 16
Discussion: https://postgr.es/m/1df00a66-db5a-4e66-809a-99b386a06d86%40enterprisedb.com
Properly update the number of bits set in the bitmap after merging the
filters in brin_bloom_union.
This is mostly harmless, as the counter is used only in the output
function, which means pageinspect may show incorrect information about
the BRIN summary. The counter does not affect correctness.
Discovered while adding a regression test comparing indexes built with
and without parallelism. The parallel index builds exercise the union
procedure when merging results from workers, which is otherwise very
hard to do in a test. Which is why this went unnoticed until now.
Backpatch through 14, where the BRIN bloom opclasses were introduced.
Backpatch-through: 14
Discussion: https://postgr.es/m/1df00a66-db5a-4e66-809a-99b386a06d86%40enterprisedb.com
aa5edbe379 has tweaked _hash_freeovflpage() so as the write buffer's
LSN is updated only when necessary, when REGBUF_NO_CHANGE is not used.
The replay code was not consistent with that, causing the write buffer's
LSN to be updated and its page to be marked as dirty even if the buffer
was registered in a "clean" state. This was possible for the case of a
squeeze record when there are no tuples to add to the write buffer, for
(is_prim_bucket_same_wrt && !is_prev_bucket_same_wrt).
I have performed some validation of this commit with
wal_consistency_checking and a change in WAL that logs REGBUF_NO_CHANGE
to a new BKPIMAGE_*. Thanks to that, it is possible to know at replay
if a buffer was clean when it was registered, then cross-checked the LSN
of the "clean" page copy coming from WAL with the LSN of the block once
the record has been replayed. This eats one bit in bimg_info, which is
not acceptable to be integrated as-is, but it could become handy in the
future. I didn't spot other areas than the one fixed by this commit at
the extent of what the main regression test suite covers.
As this is an oversight in aa5edbe379, no backpatch is required.
Reported-by: Zubeyr Eryilmaz
Author: Hayato Kuroda
Reviewed-by: Amit Kapila, Michael Paquier
Discussion: https://postgr.es/m/ZbyVVG_7eW3YD5-A@paquier.xyz
This is a C11 feature, and we require C99. While at it, go the further
step and get rid of the surrounding union (with uintptr_t) entirely,
as there is currently no use case for this file to access the header of
BlocktableEntry as a uintptr_t, and there are no additional alignment
requirements. The least invasive way seems to be to transfer the old
union name to this struct.
Reported by Pavel Borisov and Andres Freund, per buildfarm member mylodon
Reviewed by Pavel Borisov
Discussion: https://postgr.es/m/CALT9ZEH11NYV8AOzKb1bWhCf6J0H=H31f0MgT9xX+HdqvcA1rw@mail.gmail.com
Previously, the decision to store values in leaves or within the child
pointer was made at compile time, with variable length values using
leaves by necessity. This commit allows introspecting the length of
variable length values at runtime for that decision. This requires
the ability to tell whether the last-level child pointer is actually
a value, so we use a pointer tag in the lowest level bit.
Use this in TID store. This entails adding a byte to the header to
reserve space for the tag. Commit f35bd9bf3 stores up to three offsets
within the header with no bitmap, and now the header can be embedded
as above. This reduces worst-case memory usage when TIDs are sparse.
Reviewed (in an earlier version) by Masahiko Sawada
Discussion: https://postgr.es/m/CANWCAZYw+_KAaUNruhJfE=h6WgtBKeDG32St8vBJBEY82bGVRQ@mail.gmail.com
Discussion: https://postgr.es/m/CAD21AoBci3Hujzijubomo1tdwH3XtQ9F89cTNQ4bsQijOmqnEw@mail.gmail.com
While keeping API the same, this commit provides a way for block-level table
AMs to re-use existing acquire_sample_rows() by providing custom callbacks
for getting the next block and the next tuple.
Reported-by: Andres Freund
Discussion: https://postgr.es/m/20240407214001.jgpg5q3yv33ve6y3%40awork3.anarazel.de
Reviewed-by: Pavel Borisov
Let table AM define custom reloptions for its tables. This allows specifying
AM-specific parameters by the WITH clause when creating a table.
The reloptions, which could be used outside of table AM, are now extracted
into the CommonRdOptions data structure. These options could be by decision
of table AM directly specified by a user or calculated in some way.
The new test module test_tam_options evaluates the ability to set up custom
reloptions and calculate fields of CommonRdOptions on their base.
The code may use some parts from prior work by Hao Wu.
Discussion: https://postgr.es/m/CAPpHfdurb9ycV8udYqM%3Do0sPS66PJ4RCBM1g-bBpvzUfogY0EA%40mail.gmail.com
Discussion: https://postgr.es/m/AMUA1wBBBxfc3tKRLLdU64rb.1.1683276279979.Hmail.wuhao%40hashdata.cn
Reviewed-by: Reviewed-by: Pavel Borisov, Matthias van de Meent, Jess Davis
One of the assertions was the subject of a false positive complaint from
Coverity, but none of the assertions added much, so get rid of them.
Reported-By: Tom Lane <tgl@sss.pgh.pa.us>
Discussion: https://postgr.es/m/3000247.1712537309@sss.pgh.pa.us
The ANALYZE command prefetches and reads sample blocks chosen by a
BlockSampler algorithm. Instead of calling [Prefetch|Read]Buffer() for
each block, ANALYZE now uses the streaming API introduced in b5a9b18cd0.
Author: Nazir Bilal Yavuz <byavuz81@gmail.com>
Reviewed-by: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Jakub Wartak <jakub.wartak@enterprisedb.com>
Reviewed-by: Heikki Linnakangas <hlinnaka@iki.fi>
Reviewed-by: Thomas Munro <thomas.munro@gmail.com>
Discussion: https://postgr.es/m/flat/CAN55FZ0UhXqk9v3y-zW_fp4-WCp43V8y0A72xPmLkOM%2B6M%2BmJg%40mail.gmail.com
In one multixact.c edge case, we need a mechanism to wait for one
multixact offset to be written before being allowed to read the next
one. We used to handle this case by sleeping for one millisecond and
retrying, but such sleeps have been reported as problematic in
production cases. We can avoid the problem by using a condition
variable: readers sleep on it and then every creator of multixacts
broadcasts into the CV when creation is sufficiently far along.
Author: Kyotaro Horiguchi <horikyotajntt@gmail.com>
Reviewed-by: Andrey Borodin <amborodin@acm.org>
Discussion: https://postgr.es/m/47A598F4-B4E7-4029-8FEC-A06A6C3CB4B5@yandex-team.ru
Discussion: https://postgr.es/m/20200515.090333.24867479329066911.horikyota.ntt
nbtree index scans with SAOP inequalities (but no SAOP equalities)
performed extra ORDER proc lookups for any remaining equality strategy
scan keys. This could waste cycles, and caused assertion failures.
Keeping around a separate ORDER proc is only necessary for a scan's
non-array/non-SAOP equality scan keys when the scan has at least one
other SAOP equality strategy key (a SAOP inequality shouldn't count).
To fix, replace _bt_preprocess_array_keys_final's assertion with a test
that makes the function return early when the scan has no SAOP equality
scan keys.
Oversight in commit 1b134ca5, which enhanced nbtree ScalarArrayOp
execution.
Reported-By: Alexander Lakhin <exclusion@gmail.com>
Discussion: https://postgr.es/m/0539d3d3-a402-0a49-ed5e-26429dffc4bd@gmail.com
This tracks the position of WAL that's been fully copied into WAL
buffers by all processes emitting WAL. (For some reason we call that
"WAL insertion"). This is updated using atomic monotonic advance during
WaitXLogInsertionsToFinish, which is not when the insertions actually
occur, but it's the only place where we know where have all the
insertions have completed.
This value is useful in WALReadFromBuffers, which can verify that
callers don't try to read past what has been inserted. (However, more
infrastructure is needed in order to actually use WAL after the flush
point, since it could be lost.)
The value is also useful in WaitXLogInsertionsToFinish() itself, since
we can now exit quickly when all WAL has been already inserted, without
even having to take any locks.
Until now, heapgetpage()'s loop over all tuples performed some conditional
checks for each tuple, even though condition did not change across the loop.
This commit fixes that by moving the loop into an inline function. By calling
it with different constant arguments, the compiler can generate an optimized
loop for the different conditions, at the price of two per-page checks.
For cases of all-visible tables and an isolation level other than
serializable, speedups of up to 25% have been measured.
Reviewed-by: John Naylor <johncnaylorls@gmail.com>
Reviewed-by: Zhang Mingli <zmlpostgres@gmail.com>
Tested-by: Quan Zongliang <quanzongliang@yeah.net>
Discussion: https://postgr.es/m/20230716015656.xjvemfbp5fysjiea@awork3.anarazel.de
Discussion: https://postgr.es/m/2ef7ff1b-3d18-2283-61b1-bbd25fc6c7ce@yeah.net
Commit 792752af4e added infrastructure for using AVX-512 intrinsic
functions, and this commit uses that infrastructure to optimize
visibilitymap_count(). Specificially, a new pg_popcount_masked()
function is introduced that applies a bitmask to every byte in the
buffer prior to calculating the population count, which is used to
filter out the all-visible or all-frozen bits as needed. Platforms
without AVX-512 support should also see a nice speedup due to the
reduced number of calls to a function pointer.
Co-authored-by: Ants Aasma
Discussion: https://postgr.es/m/BL1PR11MB5304097DF7EA81D04C33F3D1DCA6A%40BL1PR11MB5304.namprd11.prod.outlook.com
Commit 7c70996ebf introduced an optimization to allow bitmap
scans to operate like index-only scans by not fetching a block from the
heap if none of the underlying data is needed and the block is marked
all visible in the visibility map.
With the introduction of table AMs, a FIXME was added to this code
indicating that the skip_fetch logic should be pushed into the table
AM-specific code, as not all table AMs may use a visibility map in the
same way.
This commit resolves this FIXME for the current block. The layering
violation is still present in BitmapHeapScans's prefetching code, which
uses the visibility map to decide whether or not to prefetch a block.
However, this can be addressed independently.
Author: Melanie Plageman
Reviewed-by: Andres Freund, Heikki Linnakangas, Tomas Vondra, Mark Dilger
Discussion: https://postgr.es/m/CAAKRu_ZwCwWFeL_H3ia26bP2e7HiKLWt0ZmGXPVwPO6uXq0vaA%40mail.gmail.com
Even though waiting for replay LSN happens without explicit transaction,
AbortTransaction() is responsible for the cleanup of the shared memory if
the error is thrown in a stored procedure. So, we need to do WaitLSNCleanup()
there to clean up after some unexpected error happened while waiting for
replay LSN.
Discussion: https://postgr.es/m/202404051815.eri4u5q6oj26%40alvherre.pgsql
Author: Alvaro Herrera
Commit 9e8da0f7 taught nbtree to handle ScalarArrayOpExpr quals
natively. This works by pushing down the full context (the array keys)
to the nbtree index AM, enabling it to execute multiple primitive index
scans that the planner treats as one continuous index scan/index path.
This earlier enhancement enabled nbtree ScalarArrayOp index-only scans.
It also allowed scans with ScalarArrayOp quals to return ordered results
(with some notable restrictions, described further down).
Take this general approach a lot further: teach nbtree SAOP index scans
to decide how to execute ScalarArrayOp scans (when and where to start
the next primitive index scan) based on physical index characteristics.
This can be far more efficient. All SAOP scans will now reliably avoid
duplicative leaf page accesses (just like any other nbtree index scan).
SAOP scans whose array keys are naturally clustered together now require
far fewer index descents, since we'll reliably avoid starting a new
primitive scan just to get to a later offset from the same leaf page.
The scan's arrays now advance using binary searches for the array
element that best matches the next tuple's attribute value. Required
scan key arrays (i.e. arrays from scan keys that can terminate the scan)
ratchet forward in lockstep with the index scan. Non-required arrays
(i.e. arrays from scan keys that can only exclude non-matching tuples)
"advance" without the process ever rolling over to a higher-order array.
Naturally, only required SAOP scan keys trigger skipping over leaf pages
(non-required arrays cannot safely end or start primitive index scans).
Consequently, even index scans of a composite index with a high-order
inequality scan key (which we'll mark required) and a low-order SAOP
scan key (which we won't mark required) now avoid repeating leaf page
accesses -- that benefit isn't limited to simpler equality-only cases.
In general, all nbtree index scans now output tuples as if they were one
continuous index scan -- even scans that mix a high-order inequality
with lower-order SAOP equalities reliably output tuples in index order.
This allows us to remove a couple of special cases that were applied
when building index paths with SAOP clauses during planning.
Bugfix commit 807a40c5 taught the planner to avoid generating unsafe
path keys: path keys on a multicolumn index path, with a SAOP clause on
any attribute beyond the first/most significant attribute. These cases
are now all safe, so we go back to generating path keys without regard
for the presence of SAOP clauses (just like with any other clause type).
Affected queries can now exploit scan output order in all the usual ways
(e.g., certain "ORDER BY ... LIMIT n" queries can now terminate early).
Also undo changes from follow-up bugfix commit a4523c5a, which taught
the planner to produce alternative index paths, with path keys, but
without low-order SAOP index quals (filter quals were used instead).
We'll no longer generate these alternative paths, since they can no
longer offer any meaningful advantages over standard index qual paths.
Affected queries thereby avoid all of the disadvantages that come from
using filter quals within index scan nodes. They can avoid extra heap
page accesses from using filter quals to exclude non-matching tuples
(index quals will never have that problem). They can also skip over
irrelevant sections of the index in more cases (though only when nbtree
determines that starting another primitive scan actually makes sense).
There is a theoretical risk that removing restrictions on SAOP index
paths from the planner will break compatibility with amcanorder-based
index AMs maintained as extensions. Such an index AM could have the
same limitations around ordered SAOP scans as nbtree had up until now.
Adding a pro forma incompatibility item about the issue to the Postgres
17 release notes seems like a good idea.
Author: Peter Geoghegan <pg@bowt.ie>
Author: Matthias van de Meent <boekewurm+postgres@gmail.com>
Reviewed-By: Heikki Linnakangas <hlinnaka@iki.fi>
Reviewed-By: Matthias van de Meent <boekewurm+postgres@gmail.com>
Reviewed-By: Tomas Vondra <tomas.vondra@enterprisedb.com>
Discussion: https://postgr.es/m/CAH2-Wz=ksvN_sjcnD1+Bt-WtifRA5ok48aDYnq3pkKhxgMQpcw@mail.gmail.com
This removes the need to hold both the info_lck spinlock and
WALWriteLock to update them. We use stock atomic write instead, with
WALWriteLock held. Readers can use atomic read, without any locking.
This allows for some code to be reordered: some places were a bit
contorted to avoid repeated spinlock acquisition, but that's no longer a
concern, so we can turn them to more natural coding. Some further
changes are possible (maybe to performance wins), but in this commit I
did rather minimal ones only, to avoid increasing the blast radius.
Reviewed-by: Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>
Reviewed-by: Jeff Davis <pgsql@j-davis.com>
Reviewed-by: Andres Freund <andres@anarazel.de> (earlier versions)
Discussion: https://postgr.es/m/20200831182156.GA3983@alvherre.pgsql
To allow the use of the read stream API added in b5a9b18cd for
sequential scans on heap tables, here we make some adjustments to make
that change less invasive and perhaps make the code easier to follow in
the process.
Here heapgetpage() gets broken into two functions:
1) The part which reads the block has now been moved into a function
named heapfetchbuf().
2) The part which performed pruning and populated the scan's
rs_vistuples[] array is now moved into a new function named
heap_prepare_pagescan().
The functionality provided by heap_prepare_pagescan() was only ever
required by SO_ALLOW_PAGEMODE scans, so the branching that was
previously done in heapgetpage() is no longer needed as we simply just
don't call heap_prepare_pagescan() from heapgettup() in the refactored
code.
Author: Melanie Plageman
Discussion: https://postgr.es/m/CAAKRu_YtXJiYKQvb5JsA2SkwrsizYLugs4sSOZh3EAjKUg=gEQ@mail.gmail.com
EXPLAIN (ANALYZE, SERIALIZE) allows collection of statistics about
the volume of data emitted by a query, as well as the time taken
to convert the data to the on-the-wire format. Previously there
was no way to investigate this without actually sending the data
to the client, in which case network transmission costs might
swamp what you wanted to see. In particular this feature allows
investigating the costs of de-TOASTing compressed or out-of-line
data during formatting.
Stepan Rutz and Matthias van de Meent,
reviewed by Tomas Vondra and myself
Discussion: https://postgr.es/m/ca0adb0e-fa4e-c37e-1cd7-91170b18cae1@gmx.de
After this change we have XLogCtl->logWriteResult and ->logFlushResult.
There's no functional change, other than the fact that the assignment
from shared memory to local is no longer done via struct assignment, but
instead using a macro that copies each member separately.
The current representation is inconvenient going forward; notably, we
would like to add a new member "Copy" (to keep track of the last
position copied into WAL buffers), so the symmetry between the values in
shared memory vs. those in local would be lost.
This also gives us freedom to later change the concurrency model for the
values in shared memory: we can make them use atomics instead of relying
on the info_lck spinlock.
Reviewed-by: Bharath Rupireddy <bharath.rupireddyforpostgres@gmail.com>
Discussion: https://postgr.es/m/202404031119.cd2kugjk2vho@alvherre.pgsql
Execute both freezing and pruning of tuples in the same
heap_page_prune() function, now called heap_page_prune_and_freeze(),
and emit a single WAL record containing all changes. That reduces the
overall amount of WAL generated.
This moves the freezing logic from vacuumlazy.c to the
heap_page_prune_and_freeze() function. The main difference in the
coding is that in vacuumlazy.c, we looked at the tuples after the
pruning had already happened, but in heap_page_prune_and_freeze() we
operate on the tuples before pruning. The heap_prepare_freeze_tuple()
function is now invoked after we have determined that a tuple is not
going to be pruned away.
VACUUM no longer needs to loop through the items on the page after
pruning. heap_page_prune_and_freeze() does all the work. It now
returns the list of dead offsets, including existing LP_DEAD items, to
the caller. Similarly it's now responsible for tracking 'all_visible',
'all_frozen', and 'hastup' on the caller's behalf.
Author: Melanie Plageman <melanieplageman@gmail.com>
Discussion: https://www.postgresql.org/message-id/20240330055710.kqg6ii2cdojsxgje@liskov
In preparation of freezing and counting tuples which are not
candidates for pruning, split heap_prune_record_unchanged() into
multiple functions, depending the kind of line pointer. That's not too
interesting right now, but makes the next commit smaller.
Recording the lowest soon-to-be prunable xid is one of the actions we
take for unchanged LP_NORMAL item pointers but not for others, so move
that to the new heap_prune_record_unchanged_lp_normal() function. The
next commit will add more actions to these functions.
Author: Melanie Plageman <melanieplageman@gmail.com>
Discussion: https://www.postgresql.org/message-id/20240330055710.kqg6ii2cdojsxgje@liskov
06c418e163 introduced pg_wal_replay_wait() procedure allowing to wait for
the particular LSN to be replayed on standby. The waiters were stored in
the flat array. Even though scanning small arrays is fast, that might be a
problem at scale (a lot of waiting processes).
This commit replaces the flat shared memory array with the pairing heap,
which holds the waiter with the least LSN at the top. This gives us O(log N)
complexity for both inserting and removing waiters.
Reported-by: Alvaro Herrera
Discussion: https://postgr.es/m/202404030658.hhj3vfxeyhft%40alvherre.pgsql
This adds errcodes to a set of PANIC and FATAL errors in xlog.c
and relcache.c, which previously had no errcode at all set, in
order to make fleetwide analysis of errorlogs easier. There are
many more ereport/elogs left which could benefit from having an
errcode but this at least makes a dent in the issue.
Author: Nazir Bilal Yavuz <byavuz81@gmail.com>
Discussion: https://postgr.es/m/CAN55FZ1k8LgLEqncPGmz_fWnrobV6bjABOTH4tOWta6xNcPQig@mail.gmail.com
pg_wal_replay_wait() is to be used on standby and specifies waiting for
the specific WAL location to be replayed before starting the transaction.
This option is useful when the user makes some data changes on primary and
needs a guarantee to see these changes on standby.
The queue of waiters is stored in the shared memory array sorted by LSN.
During replay of WAL waiters whose LSNs are already replayed are deleted from
the shared memory array and woken up by setting of their latches.
pg_wal_replay_wait() needs to wait without any snapshot held. Otherwise,
the snapshot could prevent the replay of WAL records implying a kind of
self-deadlock. This is why it is only possible to implement
pg_wal_replay_wait() as a procedure working in a non-atomic context,
not a function.
Catversion is bumped.
Discussion: https://postgr.es/m/eb12f9b03851bb2583adab5df9579b4b%40postgrespro.ru
Author: Kartyshov Ivan, Alexander Korotkov
Reviewed-by: Michael Paquier, Peter Eisentraut, Dilip Kumar, Amit Kapila
Reviewed-by: Alexander Lakhin, Bharath Rupireddy, Euler Taveira
Previously, we used a simple array for storing dead tuple IDs during
lazy vacuum, which had a number of problems:
* The array used a single allocation and so was limited to 1GB.
* The allocation was pessimistically sized according to table size.
* Lookup with binary search was slow because of poor CPU cache and
branch prediction behavior.
This commit replaces that array with the TID store from commit
30e144287a.
Since the backing radix tree makes small allocations as needed, the
1GB limit is now gone. Further, the total memory used is now often
smaller by an order of magnitude or more, depending on the
distribution of blocks and offsets. These two features should make
multiple rounds of heap scanning and index cleanup an extremely rare
event. TID lookup during index cleanup is also several times faster,
even more so when index order is correlated with heap tuple order.
Since there is no longer a predictable relationship between the number
of dead tuples vacuumed and the space taken up by their TIDs, the
number of tuples no longer provides any meaningful insights for users,
nor is the maximum number predictable. For that reason this commit
also changes to byte-based progress reporting, with the relevant
columns of pg_stat_progress_vacuum renamed accordingly to
max_dead_tuple_bytes and dead_tuple_bytes.
For parallel vacuum, both the TID store and supplemental information
specific to vacuum are shared among the parallel vacuum workers. As
with the previous array, we don't take any locks on TidStore during
parallel vacuum since writes are still only done by the leader
process.
Bump catalog version.
Reviewed-by: John Naylor, (in an earlier version) Dilip Kumar
Discussion: https://postgr.es/m/CAD21AoAfOZvmfR0j8VmZorZjL7RhTiQdVttNuC4W-Shdc2a-AA%40mail.gmail.com
Peter Geoghegan
Tomas Vondra
Daniel Gustafsson
Alexander Korotkov
Michael Paquier
John Naylor
Andres Freund
Thomas Munro
Alvaro Herrera
Nathan Bossart
David Rowley
Tom Lane
Heikki Linnakangas
Masahiko Sawada