Adjust some "can't happen" error messages that assumed that the page
deletion target page must be a half-dead page. This assumption was
wrong in the case of an internal target page. Simply refer to these
pages as the target page instead.
Internal pages are never marked half-dead. There is exactly one
half-dead page for each subtree undergoing deletion. The half-dead page
is also the target subtree's leaf-level page. This has been the case
since commit efada2b8, which totally overhauled nbtree page deletion.
Teach VACUUM VERBOSE to report on pages deleted by the _current_ VACUUM
operation -- these are newly deleted pages. VACUUM VERBOSE continues to
report on the total number of deleted pages in the entire index (no
change there). The former is a subset of the latter.
The distinction between each category of deleted index page only arises
with index AMs where page deletion is supported and is decoupled from
page recycling for performance reasons.
This is follow-up work to commit e5d8a999, which made nbtree store
64-bit XIDs (not 32-bit XIDs) in pages at the point at which they're
deleted. Note that the btm_last_cleanup_num_delpages metapage field
added by that commit usually gets set to pages_newly_deleted. The
exceptions (the scenarios in which they're not equal) all seem to be
tricky cases for the implementation (of page deletion and recycling) in
general.
Author: Peter Geoghegan <pg@bowt.ie>
Discussion: https://postgr.es/m/CAH2-WznpdHvujGUwYZ8sihX%3Dd5u-tRYhi-F4wnV2uN2zHpMUXw%40mail.gmail.com
Otherwise we risk "leaking" deleted pages by making them non-recyclable
indefinitely. Commit 6655a729 did the same thing for deleted pages in
GiST indexes. That work was used as a starting point here.
Stop storing an XID indicating the oldest bpto.xact across all deleted
though unrecycled pages in nbtree metapages. There is no longer any
reason to care about that condition/the oldest XID. It only ever made
sense when wraparound was something _bt_vacuum_needs_cleanup() had to
consider.
The btm_oldest_btpo_xact metapage field has been repurposed and renamed.
It is now btm_last_cleanup_num_delpages, which is used to remember how
many non-recycled deleted pages remain from the last VACUUM (in practice
its value is usually the precise number of pages that were _newly
deleted_ during the specific VACUUM operation that last set the field).
The general idea behind storing btm_last_cleanup_num_delpages is to use
it to give _some_ consideration to non-recycled deleted pages inside
_bt_vacuum_needs_cleanup() -- though never too much. We only really
need to avoid leaving a truly excessive number of deleted pages in an
unrecycled state forever. We only do this to cover certain narrow cases
where no other factor makes VACUUM do a full scan, and yet the index
continues to grow (and so actually misses out on recycling existing
deleted pages).
These metapage changes result in a clear user-visible benefit: We no
longer trigger full index scans during VACUUM operations solely due to
the presence of only 1 or 2 known deleted (though unrecycled) blocks
from a very large index. All that matters now is keeping the costs and
benefits in balance over time.
Fix an issue that has been around since commit 857f9c36, which added the
"skip full scan of index" mechanism (i.e. the _bt_vacuum_needs_cleanup()
logic). The accuracy of btm_last_cleanup_num_heap_tuples accidentally
hinged upon _when_ the source value gets stored. We now always store
btm_last_cleanup_num_heap_tuples in btvacuumcleanup(). This fixes the
issue because IndexVacuumInfo.num_heap_tuples (the source field) is
expected to accurately indicate the state of the table _after_ the
VACUUM completes inside btvacuumcleanup().
A backpatchable fix cannot easily be extracted from this commit. A
targeted fix for the issue will follow in a later commit, though that
won't happen today.
I (pgeoghegan) have chosen to remove any mention of deleted pages in the
documentation of the vacuum_cleanup_index_scale_factor GUC/param, since
the presence of deleted (though unrecycled) pages is no longer of much
concern to users. The vacuum_cleanup_index_scale_factor description in
the docs now seems rather unclear in any case, and it should probably be
rewritten in the near future. Perhaps some passing mention of page
deletion will be added back at the same time.
Bump XLOG_PAGE_MAGIC due to nbtree WAL records using full XIDs now.
Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: Masahiko Sawada <sawada.mshk@gmail.com>
Discussion: https://postgr.es/m/CAH2-WznpdHvujGUwYZ8sihX=d5u-tRYhi-F4wnV2uN2zHpMUXw@mail.gmail.com
Add some additional defensive checks in the second phase of index
deletion to detect and report index corruption during VACUUM, and to
avoid having VACUUM become stuck in more cases. The code is still not
robust in the presence of a circular chain of sibling links, though it's
not clear whether that really matters. This is follow-up work to commit
3a01f68e.
The new defensive checks rely on the assumption that there can be no
more than one VACUUM operation running for an index at any given time.
Remove an old comment suggesting that multiple concurrent VACUUMs need
to be considered here. This concern now seems highly unlikely to have
any real validity, since we clearly rely on the same assumption in
several other places. For example, there are much more recent comments
that appear in the same function (added by commit efada2b8e9) that make
the same assumption.
Also add a CHECK_FOR_INTERRUPTS() to the relevant code path. Contrary
to comments added by commit 3a01f68e, it is actually possible to handle
interrupts here, at least in the common case where processing takes
place at the leaf level. We only hold a pin on leafbuf/target page when
stepping right at the leaf level.
No backpatch due to the lack of complaints following hardening added to
the same area by commit 3a01f68e.
The latestRemovedXid values used by nbtree deletion operations are
determined by _bt_delitems_delete()'s caller, so there is no reason to
pass a separate heapRel argument.
Oversight in commit d168b66682.
Teach nbtree and heapam to cooperate in order to eagerly remove
duplicate tuples representing dead MVCC versions. This is "bottom-up
deletion". Each bottom-up deletion pass is triggered lazily in response
to a flood of versions on an nbtree leaf page. This usually involves a
"logically unchanged index" hint (these are produced by the executor
mechanism added by commit 9dc718bd).
The immediate goal of bottom-up index deletion is to avoid "unnecessary"
page splits caused entirely by version duplicates. It naturally has an
even more useful effect, though: it acts as a backstop against
accumulating an excessive number of index tuple versions for any given
_logical row_. Bottom-up index deletion complements what we might now
call "top-down index deletion": index vacuuming performed by VACUUM.
Bottom-up index deletion responds to the immediate local needs of
queries, while leaving it up to autovacuum to perform infrequent clean
sweeps of the index. The overall effect is to avoid certain
pathological performance issues related to "version churn" from UPDATEs.
The previous tableam interface used by index AMs to perform tuple
deletion (the table_compute_xid_horizon_for_tuples() function) has been
replaced with a new interface that supports certain new requirements.
Many (perhaps all) of the capabilities added to nbtree by this commit
could also be extended to other index AMs. That is left as work for a
later commit.
Extend deletion of LP_DEAD-marked index tuples in nbtree by adding logic
to consider extra index tuples (that are not LP_DEAD-marked) for
deletion in passing. This increases the number of index tuples deleted
significantly in many cases. The LP_DEAD deletion process (which is now
called "simple deletion" to clearly distinguish it from bottom-up
deletion) won't usually need to visit any extra table blocks to check
these extra tuples. We have to visit the same table blocks anyway to
generate a latestRemovedXid value (at least in the common case where the
index deletion operation's WAL record needs such a value).
Testing has shown that the "extra tuples" simple deletion enhancement
increases the number of index tuples deleted with almost any workload
that has LP_DEAD bits set in leaf pages. That is, it almost never fails
to delete at least a few extra index tuples. It helps most of all in
cases that happen to naturally have a lot of delete-safe tuples. It's
not uncommon for an individual deletion operation to end up deleting an
order of magnitude more index tuples compared to the old naive approach
(e.g., custom instrumentation of the patch shows that this happens
fairly often when the regression tests are run).
Add a further enhancement that augments simple deletion and bottom-up
deletion in indexes that make use of deduplication: Teach nbtree's
_bt_delitems_delete() function to support granular TID deletion in
posting list tuples. It is now possible to delete individual TIDs from
posting list tuples provided the TIDs have a tableam block number of a
table block that gets visited as part of the deletion process (visiting
the table block can be triggered directly or indirectly). Setting the
LP_DEAD bit of a posting list tuple is still an all-or-nothing thing,
but that matters much less now that deletion only needs to start out
with the right _general_ idea about which index tuples are deletable.
Bump XLOG_PAGE_MAGIC because xl_btree_delete changed.
No bump in BTREE_VERSION, since there are no changes to the on-disk
representation of nbtree indexes. Indexes built on PostgreSQL 12 or
PostgreSQL 13 will automatically benefit from bottom-up index deletion
(i.e. no reindexing required) following a pg_upgrade. The enhancement
to simple deletion is available with all B-Tree indexes following a
pg_upgrade, no matter what PostgreSQL version the user upgrades from.
Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: Heikki Linnakangas <hlinnaka@iki.fi>
Reviewed-By: Victor Yegorov <vyegorov@gmail.com>
Discussion: https://postgr.es/m/CAH2-Wzm+maE3apHB8NOtmM=p-DO65j2V5GzAWCOEEuy3JZgb2g@mail.gmail.com
User-visible log messages should go through ereport(), so they are
subject to translation. Many remaining elog(LOG) calls are really
debugging calls.
Reviewed-by: Alvaro Herrera <alvherre@alvh.no-ip.org>
Reviewed-by: Michael Paquier <michael@paquier.xyz>
Reviewed-by: Noah Misch <noah@leadboat.com>
Discussion: https://www.postgresql.org/message-id/flat/92d6f545-5102-65d8-3c87-489f71ea0a37%40enterprisedb.com
Streamline handling of the various strategies that we have to avoid a
page split in nbtinsert.c. When it looks like a leaf page is about to
overflow, we now perform deleting LP_DEAD items and deduplication in one
central place. This greatly simplifies _bt_findinsertloc().
This has an independently useful consequence: nbtree no longer relies on
the BTP_HAS_GARBAGE page level flag/hint for anything important. We
still set and unset the flag in the same way as before, but it's no
longer treated as a gating condition when considering if we should check
for already-set LP_DEAD bits. This happens at the point where the page
looks like it might have to be split anyway, so simply checking the
LP_DEAD bits in passing is practically free. This avoids missing
LP_DEAD bits just because the page-level hint is unset, which is
probably reasonably common (e.g. it happens when VACUUM unsets the
page-level flag without actually removing index tuples whose LP_DEAD-bit
was set recently, after the VACUUM operation began but before it reached
the leaf page in question).
Note that this isn't a big behavioral change compared to PostgreSQL 13.
We were already checking for set LP_DEAD bits regardless of whether the
BTP_HAS_GARBAGE page level flag was set before we considered doing a
deduplication pass. This commit only goes slightly further by doing the
same check for all indexes, even indexes where deduplication won't be
performed.
We don't completely remove the BTP_HAS_GARBAGE flag. We still rely on
it as a gating condition with pg_upgrade'd indexes from before B-tree
version 4/PostgreSQL 12. That makes sense because we sometimes have to
make a choice among pages full of duplicates when inserting a tuple with
pre version 4 indexes. It probably still pays to avoid accessing the
line pointer array of a page there, since it won't yet be clear whether
we'll insert on to the page in question at all, let alone split it as a
result.
Author: Peter Geoghegan <pg@bowt.ie>
Reviewed-By: Victor Yegorov <vyegorov@gmail.com>
Discussion: https://postgr.es/m/CAH2-Wz%3DYpc1PDdk8OVJDChGJBjT06%3DA0Mbv9HyTLCsOknGcUFg%40mail.gmail.com
Now that xmin isn't needed for GetSnapshotData() anymore, it leads to
unnecessary cacheline ping-pong to have it in PGXACT, as it is updated
considerably more frequently than the other PGXACT members.
After the changes in dc7420c2c9, this is a very straight-forward change.
For highly concurrent, snapshot acquisition heavy, workloads this change alone
can significantly increase scalability. E.g. plain pgbench on a smaller 2
socket machine gains 1.07x for read-only pgbench, 1.22x for read-only pgbench
when submitting queries in batches of 100, and 2.85x for batches of 100
'SELECT';. The latter numbers are obviously not to be expected in the
real-world, but micro-benchmark the snapshot computation
scalability (previously spending ~80% of the time in GetSnapshotData()).
Author: Andres Freund <andres@anarazel.de>
Reviewed-By: Robert Haas <robertmhaas@gmail.com>
Reviewed-By: Thomas Munro <thomas.munro@gmail.com>
Reviewed-By: David Rowley <dgrowleyml@gmail.com>
Discussion: https://postgr.es/m/20200301083601.ews6hz5dduc3w2se@alap3.anarazel.de
To make GetSnapshotData() more scalable, it cannot not look at at each proc's
xmin: While snapshot contents do not need to change whenever a read-only
transaction commits or a snapshot is released, a proc's xmin is modified in
those cases. The frequency of xmin modifications leads to, particularly on
higher core count systems, many cache misses inside GetSnapshotData(), despite
the data underlying a snapshot not changing. That is the most
significant source of GetSnapshotData() scaling poorly on larger systems.
Without accessing xmins, GetSnapshotData() cannot calculate accurate horizons /
thresholds as it has so far. But we don't really have to: The horizons don't
actually change that much between GetSnapshotData() calls. Nor are the horizons
actually used every time a snapshot is built.
The trick this commit introduces is to delay computation of accurate horizons
until there use and using horizon boundaries to determine whether accurate
horizons need to be computed.
The use of RecentGlobal[Data]Xmin to decide whether a row version could be
removed has been replaces with new GlobalVisTest* functions. These use two
thresholds to determine whether a row can be pruned:
1) definitely_needed, indicating that rows deleted by XIDs >= definitely_needed
are definitely still visible.
2) maybe_needed, indicating that rows deleted by XIDs < maybe_needed can
definitely be removed
GetSnapshotData() updates definitely_needed to be the xmin of the computed
snapshot.
When testing whether a row can be removed (with GlobalVisTestIsRemovableXid())
and the tested XID falls in between the two (i.e. XID >= maybe_needed && XID <
definitely_needed) the boundaries can be recomputed to be more accurate. As it
is not cheap to compute accurate boundaries, we limit the number of times that
happens in short succession. As the boundaries used by
GlobalVisTestIsRemovableXid() are never reset (with maybe_needed updated by
GetSnapshotData()), it is likely that further test can benefit from an earlier
computation of accurate horizons.
To avoid regressing performance when old_snapshot_threshold is set (as that
requires an accurate horizon to be computed), heap_page_prune_opt() doesn't
unconditionally call TransactionIdLimitedForOldSnapshots() anymore. Both the
computation of the limited horizon, and the triggering of errors (with
SetOldSnapshotThresholdTimestamp()) is now only done when necessary to remove
tuples.
This commit just removes the accesses to PGXACT->xmin from
GetSnapshotData(), but other members of PGXACT residing in the same
cache line are accessed. Therefore this in itself does not result in a
significant improvement. Subsequent commits will take advantage of the
fact that GetSnapshotData() now does not need to access xmins anymore.
Note: This contains a workaround in heap_page_prune_opt() to keep the
snapshot_too_old tests working. While that workaround is ugly, the tests
currently are not meaningful, and it seems best to address them separately.
Author: Andres Freund <andres@anarazel.de>
Reviewed-By: Robert Haas <robertmhaas@gmail.com>
Reviewed-By: Thomas Munro <thomas.munro@gmail.com>
Reviewed-By: David Rowley <dgrowleyml@gmail.com>
Discussion: https://postgr.es/m/20200301083601.ews6hz5dduc3w2se@alap3.anarazel.de
Currently, page unlink leaves remaining items "as is", but replay of
corresponding WAL-record re-initializes page leaving it with no items.
For the sake of consistency, this commit makes primary delete all the items
during page unlink as well.
Thanks to this change, we now don't mask contents of deleted btree page for
WAL consistency checking.
Discussion: https://postgr.es/m/CAPpHfdt_OTyQpXaPJcWzV2N-LNeNJseNB-K_A66qG%3DL518VTFw%40mail.gmail.com
Author: Alexander Korotkov
Reviewed-by: Peter Geoghegan
Add a documenting assertion that's similar to the nearby assertion added
by commit cd8c73a3. This conveys that the entire call to _bt_pagedel()
does no work if it isn't possible to get a descent stack for the initial
scanblkno page.
Holding just a buffer pin (with no buffer lock) on an nbtree buffer/page
provides very weak guarantees, especially compared to heapam, where it's
often safe to read a page while only holding a buffer pin. This commit
has Valgrind enforce the following rule: it is never okay to access an
nbtree buffer without holding both a pin and a lock on the buffer.
A draft version of this patch detected questionable code that was
cleaned up by commits fa7ff642 and 7154aa16. The code in question used
to access an nbtree buffer page's special/opaque area with no buffer
lock (only a buffer pin). This practice (which isn't obviously unsafe)
is hereby formally disallowed in nbtree. There doesn't seem to be any
reason to allow it, and banning it keeps things simple for Valgrind.
The new checks are implemented by adding custom nbtree client requests
(located in LockBuffer() wrapper functions); these requests are
"superimposed" on top of the generic bufmgr.c Valgrind client requests
added by commit 1e0dfd16. No custom resource management cleanup code is
needed to undo the effects of marking buffers as non-accessible under
this scheme.
Author: Peter Geoghegan
Reviewed-By: Anastasia Lubennikova, Georgios Kokolatos
Discussion: https://postgr.es/m/CAH2-WzkLgyN3zBvRZ1pkNJThC=xi_0gpWRUb_45eexLH1+k2_Q@mail.gmail.com
Commit 624686abcf added an assertion that verified that _bt_search
successfully relocated the leaf page undergoing deletion. Page deletion
cannot deal with the case where the descent stack is to the right of the
page, so this seemed critical (deletion can only handle the case where
the descent stack is to the left of the leaf/target page). However, the
assertion went a bit too far.
Since only a buffer pin is held on the leaf page throughout the call to
_bt_search, nothing guarantees that it can't have split during this
small window. And if does actually split, _bt_search may end up
"relocating" a page to the right of the original target leaf page. This
scenario seems extremely unlikely, but it must still be considered.
Remove the assertion, and document how we cope in this scenario.
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.
Restructure the function that locates the root of the to-be-deleted
subtree during nbtree page deletion. Handle the conditions that make
page deletion unsafe in a slightly more uniform way, and acknowledge the
fact that the behavior with incomplete splits on internal pages is
different (as pointed out in the nbtree README as of commit 35bc0ec7).
Also invent new terminology that avoids ambiguity around which pages are
about to be deleted. Consistently use the term "to-be-deleted subtree",
not the ambiguous term "branch".
We were calling the subtree parent page the "top parent page", but that
was quite misleading. The top parent page usually refers to a page
unlinked from its siblings and marked deleted (during the second stage
of page deletion). There was one kind of top parent page that we merely
removed a downlink from, and another kind of top parent page that we
actually marked deleted. Eliminate the ambiguity by inventing a new
term ("subtree parent page") that refers to the former kind of page
only.
Factor out code common to _bt_lock_branch_parent() and _bt_pagedel()
into a new utility function. This new function is used to check that
the left sibling of a deletion target page does not have the
INCOMPLETE_SPLIT page flag set. If it is set then deletion is unsafe;
there won't be a usable pivot tuple (with a downlink) in the parent page
that points to the deletion target page. The page deletion algorithm is
not prepared to deal with that. Also restructure an existing, related
utility function that checks if the right sibling of the target page has
the ISHALFDEAD page flag set.
This organization highlights the symmetry between the two cases. The
goal is to make the design of page deletion clearer. Both functions
involve a sibling page with a flag that indicates that there was an
interrupted operation (a page split or a page deletion) that resulted in
a page pointed to by sibling pages, but not pointed to in the parent.
And, both functions indicate if page deletion is unsafe due to the
absence of a particular downlink in the parent page.
The logic for determining how many nbtree pages in an index are deleted
pages sometimes undercounted pages. Pages that were deleted by the
current VACUUM operation (as opposed to some previous VACUUM operation
whose deleted pages have yet to be reused) were sometimes overlooked.
The final count is exposed to users through VACUUM VERBOSE's "%u index
pages have been deleted" output.
btvacuumpage() avoided double-counting when _bt_pagedel() deleted more
than one page by assuming that only one page was deleted, and that the
additional deleted pages would get picked up during a future call to
btvacuumpage() by the same VACUUM operation. _bt_pagedel() can
legitimately delete pages that the btvacuumscan() scan will not visit
again, though, so that assumption was slightly faulty.
Fix the accounting by teaching _bt_pagedel() about its caller's
requirements. It now only reports on pages that it knows btvacuumscan()
won't visit again (including the current btvacuumpage() page), so
everything works out in the end.
This bug has been around forever. Only backpatch to v11, though, to
keep _bt_pagedel() is sync on the branches that have today's bugfix
commit b0229f26da. Note that this commit changes the signature of
_bt_pagedel(), just like commit b0229f26da.
Author: Peter Geoghegan
Reviewed-By: Masahiko Sawada
Discussion: https://postgr.es/m/CAH2-WzkrXBcMQWAYUJMFTTvzx_r4q=pYSjDe07JnUXhe+OZnJA@mail.gmail.com
Backpatch: 11-
Commit 857f9c36cd (which taught nbtree VACUUM to skip a scan of the
index from btcleanup in situations where it doesn't seem worth it) made
VACUUM maintain the oldest btpo.xact among all deleted pages for the
index as a whole. It failed to handle all the details surrounding pages
that are deleted by the current VACUUM operation correctly (though pages
deleted by some previous VACUUM operation were processed correctly).
The most immediate problem was that the special area of the page was
examined without a buffer pin at one point. More fundamentally, the
handling failed to account for the full range of _bt_pagedel()
behaviors. For example, _bt_pagedel() sometimes deletes internal pages
in passing, as part of deleting an entire subtree with btvacuumpage()
caller's page as the leaf level page. The original leaf page passed to
_bt_pagedel() might not be the page that it deletes first in cases where
deletion can take place.
It's unclear how disruptive this bug may have been, or what symptoms
users might want to look out for. The issue was spotted during
unrelated code review.
To fix, push down the logic for maintaining the oldest btpo.xact to
_bt_pagedel(). btvacuumpage() is now responsible for pages that were
fully deleted by a previous VACUUM operation, while _bt_pagedel() is now
responsible for pages that were deleted by the current VACUUM operation
(this includes half-dead pages from a previous interrupted VACUUM
operation that become fully deleted in _bt_pagedel()). Note that
_bt_pagedel() should never encounter an existing deleted page.
This commit theoretically breaks the ABI of a stable release by changing
the signature of _bt_pagedel(). However, if any third party extension
is actually affected by this, then it must already be completely broken
(since there are numerous assumptions made in _bt_pagedel() that cannot
be met outside of VACUUM). It seems highly unlikely that such an
extension actually exists, in any case.
Author: Peter Geoghegan
Reviewed-By: Masahiko Sawada
Discussion: https://postgr.es/m/CAH2-WzkrXBcMQWAYUJMFTTvzx_r4q=pYSjDe07JnUXhe+OZnJA@mail.gmail.com
Backpatch: 11-, where the "skip full scan" feature was introduced.
Avoid accessing the leaf page's top parent tuple without a buffer lock
held during the second phase of nbtree page deletion. The old approach
was safe, though only because VACUUM never drops its buffer pin (and
because only VACUUM itself can modify a half-dead page). Even still, it
seems like a good idea to be strict here. Tighten things up by copying
the top parent page's block number to a local variable before releasing
the buffer lock on the leaf page -- not after.
This is a follow-up to commit fa7ff642, which fixed a similar issue in
the first phase of nbtree page deletion.
Update some related comments in passing.
Discussion: https://postgr.es/m/CAH2-WzkLgyN3zBvRZ1pkNJThC=xi_0gpWRUb_45eexLH1+k2_Q@mail.gmail.com
Avoid accessing the deletion target page's special area during nbtree
page deletion at a point where there is no buffer lock held. This issue
was detected by a patch that extends Valgrind's memcheck tool to mark
nbtree pages that are unsafe to access (due to not having a buffer lock
or buffer pin) as NOACCESS.
We do hold a buffer pin at this point, and only access the special area,
so the old approach was safe. Even still, it seems like a good idea to
tighten up the rules in this area. There is no reason to not simply
insist on always holding a buffer lock (not just a pin) when accessing
nbtree pages.
Update some related comments in passing.
Discussion: https://postgr.es/m/CAH2-WzkLgyN3zBvRZ1pkNJThC=xi_0gpWRUb_45eexLH1+k2_Q@mail.gmail.com
Deduplication reduces the storage overhead of duplicates in indexes that
use the standard nbtree index access method. The deduplication process
is applied lazily, after the point where opportunistic deletion of
LP_DEAD-marked index tuples occurs. Deduplication is only applied at
the point where a leaf page split would otherwise be required. New
posting list tuples are formed by merging together existing duplicate
tuples. The physical representation of the items on an nbtree leaf page
is made more space efficient by deduplication, but the logical contents
of the page are not changed. Even unique indexes make use of
deduplication as a way of controlling bloat from duplicates whose TIDs
point to different versions of the same logical table row.
The lazy approach taken by nbtree has significant advantages over a GIN
style eager approach. Most individual inserts of index tuples have
exactly the same overhead as before. The extra overhead of
deduplication is amortized across insertions, just like the overhead of
page splits. The key space of indexes works in the same way as it has
since commit dd299df8 (the commit that made heap TID a tiebreaker
column).
Testing has shown that nbtree deduplication can generally make indexes
with about 10 or 15 tuples for each distinct key value about 2.5X - 4X
smaller, even with single column integer indexes (e.g., an index on a
referencing column that accompanies a foreign key). The final size of
single column nbtree indexes comes close to the final size of a similar
contrib/btree_gin index, at least in cases where GIN's posting list
compression isn't very effective. This can significantly improve
transaction throughput, and significantly reduce the cost of vacuuming
indexes.
A new index storage parameter (deduplicate_items) controls the use of
deduplication. The default setting is 'on', so all new B-Tree indexes
automatically use deduplication where possible. This decision will be
reviewed at the end of the Postgres 13 beta period.
There is a regression of approximately 2% of transaction throughput with
synthetic workloads that consist of append-only inserts into a table
with several non-unique indexes, where all indexes have few or no
repeated values. The underlying issue is that cycles are wasted on
unsuccessful attempts at deduplicating items in non-unique indexes.
There doesn't seem to be a way around it short of disabling
deduplication entirely. Note that deduplication of items in unique
indexes is fairly well targeted in general, which avoids the problem
there (we can use a special heuristic to trigger deduplication passes in
unique indexes, since we're specifically targeting "version bloat").
Bump XLOG_PAGE_MAGIC because xl_btree_vacuum changed.
No bump in BTREE_VERSION, since the representation of posting list
tuples works in a way that's backwards compatible with version 4 indexes
(i.e. indexes built on PostgreSQL 12). However, users must still
REINDEX a pg_upgrade'd index to use deduplication, regardless of the
Postgres version they've upgraded from. This is the only way to set the
new nbtree metapage flag indicating that deduplication is generally
safe.
Author: Anastasia Lubennikova, Peter Geoghegan
Reviewed-By: Peter Geoghegan, Heikki Linnakangas
Discussion:
https://postgr.es/m/55E4051B.7020209@postgrespro.ruhttps://postgr.es/m/4ab6e2db-bcee-f4cf-0916-3a06e6ccbb55@postgrespro.ru
Commit 558a9165e0 taught _bt_delitems_delete() to produce its own XID
horizon on the primary. Standbys no longer needed to generate their own
latestRemovedXid, since they could just use the explicitly logged value
from the primary instead. The deleted offset numbers array from the
xl_btree_delete WAL record was no longer used by the REDO routine for
anything other than deleting the items.
This enables a minor optimization: We now treat the array as buffer
state, not generic WAL data, following _bt_delitems_vacuum()'s example.
This should be a minor win, since it allows us to avoid including the
deleted items array in cases where XLogInsert() stores the whole buffer
anyway. The primary goal here is to make the code more maintainable,
though. Removing inessential differences between the two functions
highlights the fundamental differences that remain.
Also change xl_btree_delete to use uint32 for the size of the array of
item offsets being deleted. This brings xl_btree_delete closer to
xl_btree_vacuum. Furthermore, it seems like a good idea to use an
explicit-width integer type (the field was previously an "int").
Bump XLOG_PAGE_MAGIC because xl_btree_delete changed.
Discussion: https://postgr.es/m/CAH2-Wzkz4TjmezzfAbaV1zYrh=fr0bCpzuJTvBe5iUQ3aUPsCQ@mail.gmail.com
_bt_delitems_vacuum() comments claimed that it isn't worth another scan
of the page to avoid falsely unsetting the BTP_HAS_GARBAGE page flag
hint (this happens to be the same wording that was removed from
_bt_delitems_delete() by my recent commit fe97c61c). The comments made
little sense, though. The issue can't have much to do with performing a
second scan of the target leaf page, since an LP_DEAD test could easily
be performed in the first scan of the page anyway (the scan that takes
place in btvacuumpage() caller).
Revise the explanation. It makes much more sense to frame this as an
issue about recovery conflicts. _bt_delitems_vacuum() cannot easily
generate an XID cutoff in the same way that _bt_delitems_delete() is
designed to.
Falsely unsetting the page flag is not ideal, and is likely to happen
more often than was supposed by the original comments. Explain why it
usually isn't a problem in practice. There may be an argument for
_bt_delitems_vacuum() not clearing the BTP_HAS_GARBAGE bit, removing the
question of it being falsely unset by VACUUM (there may even be an
argument for not using a page level hint at all). This can be revisited
later.
Comments about the consequences of clearing the BTP_HAS_GARBAGE page
flag bit that apply only to VACUUM were added to code that deals with
opportunistic deletion of LP_DEAD items by commit a760893d. The same
comment block was added to both _bt_delitems_vacuum() and
_bt_delitems_delete(). Correct _bt_delitems_delete()'s copy of the
comment block.
_bt_delitems_delete() reliably deletes items that were found by caller
to have their LP_DEAD bit set. There is no question about whether or
not unsetting the BTP_HAS_GARBAGE bit can miss some LP_DEAD items that
were set recently.
Also tweak a related section of the nbtree README.
The REDO routine for nbtree's xl_btree_vacuum record type hasn't
performed a "pin scan" since commit 3e4b7d87 went in, so clearly there
isn't any point in VACUUM WAL-logging information that won't actually be
used. Finish off the work of commit 3e4b7d87 (and the closely related
preceding commit 687f2cd7) by removing the code that generates this
unused information. Also remove the REDO routine code disabled by
commit 3e4b7d87.
Replace the unneeded lastBlockVacuumed field in xl_btree_vacuum with a
new "ndeleted" field. The new field isn't actually needed right now,
since we could continue to infer the array length from the overall
record length. However, an upcoming patch to add deduplication to
nbtree needs to add an "items updated" field to xl_btree_vacuum, so we
might as well start being explicit about the number of items now.
(Besides, it doesn't seem like a good idea to leave the xl_btree_vacuum
struct without any fields; the C standard says that that's undefined.)
nbtree VACUUM no longer forces writing a WAL record for the last block
in the index. Writing out a WAL record with no items for the final
block was supposed to force processing of a lastBlockVacuumed field by a
pin scan.
Bump XLOG_PAGE_MAGIC because xl_btree_vacuum changed.
Discussion: https://postgr.es/m/CAH2-WzmY_mT7UnTzFB5LBQDBkKpdV5UxP3B5bLb7uP%3D%3D6UQJRQ%40mail.gmail.com
Commit efada2b8e9, which made the nbtree page deletion algorithm more
robust, removed the concept of a half-dead internal page. Remove a
comment about half dead parent pages that was overlooked.
The initial value of the nbtree stack downlink block number field
recorded during an initial descent of the tree wasn't actually used.
Both _bt_getstackbuf() callers overwrote the value with their own value.
Remove the block number field from the stack struct, and add a child
block number argument to _bt_getstackbuf() in its place. This makes the
overall design of _bt_getstackbuf() clearer.
Author: Peter Geoghegan
Reviewed-By: Anastasia Lubennikova
Discussion: https://postgr.es/m/CAH2-Wzmx+UbXt2YNOUCZ-a04VdXU=S=OHuAuD7Z8uQq-PXTYUg@mail.gmail.com
Use the PageIndexTupleOverwrite() bufpage.c routine within nbtree
instead of deleting a tuple and re-inserting its replacement. This
makes the intent of affected code slightly clearer. It also makes
CREATE INDEX slightly faster, since there is no longer a need to shift
every leaf page's line pointer array back and forth during index builds.
Author: Peter Geoghegan, Anastasia Lubennikova
Reviewed-By: Anastasia Lubennikova
Discussion: https://postgr.es/m/CAH2-Wz=Zk=B9+Vwm376WuO7YTjFc2SSskifQm4Nme3RRRPtOSQ@mail.gmail.com
Commit 857f9c36cd, which taught nbtree VACUUM to avoid unnecessary
index scans, bumped the nbtree version number from 2 to 3, while adding
the ability for nbtree indexes to be upgraded on-the-fly. Various
assertions that assumed that an nbtree index was always on version 2 had
to be changed to accept any supported version (version 2 or 3 on
Postgres 11).
However, a few assertions were missed in the initial commit, all of
which were in code paths that cache a local copy of the metapage
metadata, where the index had been expected to be on the current version
(no longer version 2) as a generic sanity check. Rather than simply
update the assertions, follow-up commit 0a64b45152 intentionally made
the metapage caching code update the per-backend cached metadata version
without changing the on-disk version at the same time. This could even
happen when the planner needed to determine the height of a B-Tree for
costing purposes. The assertions only fail on Postgres v12 when
upgrading from v10, because they were adjusted to use the authoritative
shared memory metapage by v12's commit dd299df8.
To fix, remove the cache-only upgrade mechanism entirely, and update the
assertions themselves to accept any supported version (go back to using
the cached version in v12). The fix is almost a full revert of commit
0a64b45152 on the v11 branch.
VACUUM only considers the authoritative metapage, and never bothers with
a locally cached version, whereas everywhere else isn't interested in
the metapage fields that were added by commit 857f9c36cd. It seems
unlikely that this bug has affected any user on v11.
Reported-By: Christoph Berg
Bug: #15896
Discussion: https://postgr.es/m/15896-5b25e260fdb0b081%40postgresql.org
Backpatch: 11-, where VACUUM was taught to avoid unnecessary index scans.
It's not safe for nbtree VACUUM to attempt to delete a target page whose
right sibling is already half-dead, since that would fail the
cross-check when VACUUM attempts to re-find a downlink to the right
sibling in the parent page. Logic to prevent this from happening was
added by commit 8da3183780, which addressed a bug in the overhaul of
page deletion that went into PostgreSQL 9.4 (commit efada2b8e9).
VACUUM was made to check the right sibling page, and back off when it
happened to be half-dead already.
However, it is only truly necessary to do the right sibling check on the
leaf level, since that transitively determines if the deletion target's
parent's right sibling page is itself undergoing deletion. Remove the
internal page level check, and add a comment explaining why the leaf
level check alone suffices.
The extra check is also unnecessary due to the fact that internal pages
that are marked half-dead are generally considered corrupt. Commit
efada2b8e9 established the principle that there should never be
half-dead internal pages (internal pages pending deletion are possible,
but that status is never directly represented in the internal page).
VACUUM will complain about corruption when it encounters half-dead
internal pages, so VACUUM is bound to raise an error one way or another
when an nbtree index has a half-dead internal page (contrib/amcheck will
also report that the page is corrupt).
It's possible that a pg_upgrade'd 9.3 database will still have half-dead
internal pages, so it may seem like there is an argument for leaving the
check in place to reliably get a cleaner error message that advises the
user to REINDEX. However, leaf pages are also deleted in the first
phase of deletion prior to PostgreSQL 9.4, so I believe we won't even
attempt to re-find the parent page anyway (we won't have the fully
deleted leaf page as the right sibling of our target page, so we won't
even try to find a downlink for it).
Discussion: https://postgr.es/m/CAH2-Wzm_ntmqJjWLRyKzimFmFvk+BnVAvUpaA4s1h9Ja58woaQ@mail.gmail.com
Previously the xid horizon was only computed during WAL replay. That
had two major problems:
1) It relied on knowing what the table pointed to looks like. That was
easy enough before the introducing of tableam (we knew it had to be
heap, although some trickery around logging the heap relfilenodes
was required). But to properly handle table AMs we need
per-database catalog access to look up the AM handler, which
recovery doesn't allow.
2) Not knowing the xid horizon also makes it hard to support logical
decoding on standbys. When on a catalog table, we need to be able
to conflict with slots that have an xid horizon that's too old. But
computing the horizon by visiting the heap only works once
consistency is reached, but we always need to be able to detect
conflicts.
There's also a secondary problem, in that the current method performs
redundant work on every standby. But that's counterbalanced by
potentially computing the value when not necessary (either because
there's no standby, or because there's no connected backends).
Solve 1) and 2) by moving computation of the xid horizon to the
primary and by involving tableam in the computation of the horizon.
To address the potentially increased overhead, increase the efficiency
of the xid horizon computation for heap by sorting the tids, and
eliminating redundant buffer accesses. When prefetching is available,
additionally perform prefetching of buffers. As this is more of a
maintenance task, rather than something routinely done in every read
only query, we add an arbitrary 10 to the effective concurrency -
thereby using IO concurrency, when not globally enabled. That's
possibly not the perfect formula, but seems good enough for now.
Bumps WAL format, as latestRemovedXid is now part of the records, and
the heap's relfilenode isn't anymore.
Author: Andres Freund, Amit Khandekar, Robert Haas
Reviewed-By: Robert Haas
Discussion:
https://postgr.es/m/20181212204154.nsxf3gzqv3gesl32@alap3.anarazel.dehttps://postgr.es/m/20181214014235.dal5ogljs3bmlq44@alap3.anarazel.dehttps://postgr.es/m/20180703070645.wchpu5muyto5n647@alap3.anarazel.de
Make nbtree treat all index tuples as having a heap TID attribute.
Index searches can distinguish duplicates by heap TID, since heap TID is
always guaranteed to be unique. This general approach has numerous
benefits for performance, and is prerequisite to teaching VACUUM to
perform "retail index tuple deletion".
Naively adding a new attribute to every pivot tuple has unacceptable
overhead (it bloats internal pages), so suffix truncation of pivot
tuples is added. This will usually truncate away the "extra" heap TID
attribute from pivot tuples during a leaf page split, and may also
truncate away additional user attributes. This can increase fan-out,
especially in a multi-column index. Truncation can only occur at the
attribute granularity, which isn't particularly effective, but works
well enough for now. A future patch may add support for truncating
"within" text attributes by generating truncated key values using new
opclass infrastructure.
Only new indexes (BTREE_VERSION 4 indexes) will have insertions that
treat heap TID as a tiebreaker attribute, or will have pivot tuples
undergo suffix truncation during a leaf page split (on-disk
compatibility with versions 2 and 3 is preserved). Upgrades to version
4 cannot be performed on-the-fly, unlike upgrades from version 2 to
version 3. contrib/amcheck continues to work with version 2 and 3
indexes, while also enforcing stricter invariants when verifying version
4 indexes. These stricter invariants are the same invariants described
by "3.1.12 Sequencing" from the Lehman and Yao paper.
A later patch will enhance the logic used by nbtree to pick a split
point. This patch is likely to negatively impact performance without
smarter choices around the precise point to split leaf pages at. Making
these two mostly-distinct sets of enhancements into distinct commits
seems like it might clarify their design, even though neither commit is
particularly useful on its own.
The maximum allowed size of new tuples is reduced by an amount equal to
the space required to store an extra MAXALIGN()'d TID in a new high key
during leaf page splits. The user-facing definition of the "1/3 of a
page" restriction is already imprecise, and so does not need to be
revised. However, there should be a compatibility note in the v12
release notes.
Author: Peter Geoghegan
Reviewed-By: Heikki Linnakangas, Alexander Korotkov
Discussion: https://postgr.es/m/CAH2-WzkVb0Kom=R+88fDFb=JSxZMFvbHVC6Mn9LJ2n=X=kS-Uw@mail.gmail.com
Use dedicated struct to represent nbtree insertion scan keys. Having a
dedicated struct makes the difference between search type scankeys and
insertion scankeys a lot clearer, and simplifies the signature of
several related functions. This is based on a suggestion by Andrey
Lepikhov.
Streamline how unique index insertions cache binary search progress.
Cache the state of in-progress binary searches within _bt_check_unique()
for later instead of having callers avoid repeating the binary search in
an ad-hoc manner. This makes it easy to add a new optimization:
_bt_check_unique() now falls out of its loop immediately in the common
case where it's already clear that there couldn't possibly be a
duplicate.
The new _bt_check_unique() scheme makes it a lot easier to manage cached
binary search effort afterwards, from within _bt_findinsertloc(). This
is needed for the upcoming patch to make nbtree tuples unique by
treating heap TID as a final tiebreaker column. Unique key binary
searches need to restore lower and upper bounds. They cannot simply
continue to use the >= lower bound as the offset to insert at, because
the heap TID tiebreaker column must be used in comparisons for the
restored binary search (unlike the original _bt_check_unique() binary
search, where scankey's heap TID column must be omitted).
Author: Peter Geoghegan, Heikki Linnakangas
Reviewed-By: Heikki Linnakangas, Andrey Lepikhov
Discussion: https://postgr.es/m/CAH2-WzmE6AhUdk9NdWBf4K3HjWXZBX3+umC7mH7+WDrKcRtsOw@mail.gmail.com
_bt_getstackbuf() is called at exactly two points following commit
efada2b8e9 (one call site is concerned with page splits, while the
other is concerned with page deletion). The parent buffer returned by
_bt_getstackbuf() is write-locked in both cases. Remove the 'access'
argument and make _bt_getstackbuf() assume that callers require a
write-lock.
Commit efada2b8e9, which made the nbtree page deletion algorithm more
robust, removed _bt_getstackbuf() calls from _bt_pagedel(). It failed
to update a comment that referenced the earlier approach. Update the
comment to explain that the _bt_getstackbuf() page deletion call site
mirrors the only other remaining _bt_getstackbuf() call site, which is
reached during page splits.
Peter Geoghegan
Thomas Munro
Bruce Momjian
Peter Eisentraut
Andres Freund
Alexander Korotkov
Tom Lane