Commit 4de82f7d7 increased the WAL flush rate, mainly to increase the
likelihood that hint bits can be set quickly. More quickly set hint bits
can reduce contention around the clog et al. But unfortunately the
increased flush rate can have a significant negative performance impact,
I have measured up to a factor of ~4. The reason for this slowdown is
that if there are independent writes to the underlying devices, for
example because shared buffers is a lot smaller than the hot data set,
or because a checkpoint is ongoing, the fdatasync() calls force cache
flushes to be emitted to the storage.
This is achieved by flushing WAL only if the last flush was longer than
wal_writer_delay ago, or if more than wal_writer_flush_after (new GUC)
unflushed blocks are pending. Based on some tests the default for
wal_writer_delay is 1MB, which seems to work well both on SSD and
rotational media.
To avoid negative performance impact due to 4de82f7d7 an earlier
commit (db76b1e) made SetHintBits() more likely to succeed; preventing
performance regressions in the pgbench tests I performed.
Discussion: 20160118163908.GW10941@awork2.anarazel.de
When a write transaction commits, it must clear its XID advertised via
the ProcArray, which requires that we hold ProcArrayLock in exclusive
mode in order to prevent concurrent processes running GetSnapshotData
from seeing inconsistent results. When many processes try to commit
at once, ProcArrayLock must change hands repeatedly, with each
concurrent process trying to commit waking up to acquire the lock in
turn. To make things more efficient, when more than one backend is
trying to commit a write transaction at the same time, have just one
of them acquire ProcArrayLock in exclusive mode and clear the XIDs of
all processes in the group. Benchmarking reveals that this is much
more efficient at very high client counts.
Amit Kapila, heavily revised by me, with some review also from Pavan
Deolasee.
Transactions can now set their commit timestamp directly as they commit,
or an external transaction commit timestamp can be fed from an outside
system using the new function TransactionTreeSetCommitTsData(). This
data is crash-safe, and truncated at Xid freeze point, same as pg_clog.
This module is disabled by default because it causes a performance hit,
but can be enabled in postgresql.conf requiring only a server restart.
A new test in src/test/modules is included.
Catalog version bumped due to the new subdirectory within PGDATA and a
couple of new SQL functions.
Authors: Álvaro Herrera and Petr Jelínek
Reviewed to varying degrees by Michael Paquier, Andres Freund, Robert
Haas, Amit Kapila, Fujii Masao, Jaime Casanova, Simon Riggs, Steven
Singer, Peter Eisentraut
Each WAL record now carries information about the modified relation and
block(s) in a standardized format. That makes it easier to write tools that
need that information, like pg_rewind, prefetching the blocks to speed up
recovery, etc.
There's a whole new API for building WAL records, replacing the XLogRecData
chains used previously. The new API consists of XLogRegister* functions,
which are called for each buffer and chunk of data that is added to the
record. The new API also gives more control over when a full-page image is
written, by passing flags to the XLogRegisterBuffer function.
This also simplifies the XLogReadBufferForRedo() calls. The function can dig
the relation and block number from the WAL record, so they no longer need to
be passed as arguments.
For the convenience of redo routines, XLogReader now disects each WAL record
after reading it, copying the main data part and the per-block data into
MAXALIGNed buffers. The data chunks are not aligned within the WAL record,
but the redo routines can assume that the pointers returned by XLogRecGet*
functions are. Redo routines are now passed the XLogReaderState, which
contains the record in the already-disected format, instead of the plain
XLogRecord.
The new record format also makes the fixed size XLogRecord header smaller,
by removing the xl_len field. The length of the "main data" portion is now
stored at the end of the WAL record, and there's a separate header after
XLogRecord for it. The alignment padding at the end of XLogRecord is also
removed. This compansates for the fact that the new format would otherwise
be more bulky than the old format.
Reviewed by Andres Freund, Amit Kapila, Michael Paquier, Alvaro Herrera,
Fujii Masao.
Every redo routine uses the same idiom to determine what to do to a page:
check if there's a backup block for it, and if not read, the buffer if the
block exists, and check its LSN. Refactor that into a common function,
XLogReadBufferForRedo, making all the redo routines shorter and more
readable.
This has no user-visible effect, and makes no changes to the WAL format.
Reviewed by Andres Freund, Alvaro Herrera, Michael Paquier.
transam/README explained how B-tree incomplete splits were tracked and
fixed after recovery, as an example of handling complex actions that need
multiple WAL records, but that's not how it works anymore. Explain the new
paradigm.
Checksums are set immediately prior to flush out of shared buffers
and checked when pages are read in again. Hint bit setting will
require full page write when block is dirtied, which causes various
infrastructure changes. Extensive comments, docs and README.
WARNING message thrown if checksum fails on non-all zeroes page;
ERROR thrown but can be disabled with ignore_checksum_failure = on.
Feature enabled by an initdb option, since transition from option off
to option on is long and complex and has not yet been implemented.
Default is not to use checksums.
Checksum used is WAL CRC-32 truncated to 16-bits.
Simon Riggs, Jeff Davis, Greg Smith
Wide input and assistance from many community members. Thank you.
Remove use of PageSetTLI() from all page manipulation functions
and adjust README to indicate change in the way we make changes
to pages. Repurpose those bytes into the pd_checksum field and
explain how that works in comments about page header.
Refactoring ahead of actual feature patch which would make use
of the checksum field, arriving later.
Jeff Davis, with comments and doc changes by Simon Riggs
Direction suggested by Robert Haas; many others providing
review comments.
This patch introduces two additional lock modes for tuples: "SELECT FOR
KEY SHARE" and "SELECT FOR NO KEY UPDATE". These don't block each
other, in contrast with already existing "SELECT FOR SHARE" and "SELECT
FOR UPDATE". UPDATE commands that do not modify the values stored in
the columns that are part of the key of the tuple now grab a SELECT FOR
NO KEY UPDATE lock on the tuple, allowing them to proceed concurrently
with tuple locks of the FOR KEY SHARE variety.
Foreign key triggers now use FOR KEY SHARE instead of FOR SHARE; this
means the concurrency improvement applies to them, which is the whole
point of this patch.
The added tuple lock semantics require some rejiggering of the multixact
module, so that the locking level that each transaction is holding can
be stored alongside its Xid. Also, multixacts now need to persist
across server restarts and crashes, because they can now represent not
only tuple locks, but also tuple updates. This means we need more
careful tracking of lifetime of pg_multixact SLRU files; since they now
persist longer, we require more infrastructure to figure out when they
can be removed. pg_upgrade also needs to be careful to copy
pg_multixact files over from the old server to the new, or at least part
of multixact.c state, depending on the versions of the old and new
servers.
Tuple time qualification rules (HeapTupleSatisfies routines) need to be
careful not to consider tuples with the "is multi" infomask bit set as
being only locked; they might need to look up MultiXact values (i.e.
possibly do pg_multixact I/O) to find out the Xid that updated a tuple,
whereas they previously were assured to only use information readily
available from the tuple header. This is considered acceptable, because
the extra I/O would involve cases that would previously cause some
commands to block waiting for concurrent transactions to finish.
Another important change is the fact that locking tuples that have
previously been updated causes the future versions to be marked as
locked, too; this is essential for correctness of foreign key checks.
This causes additional WAL-logging, also (there was previously a single
WAL record for a locked tuple; now there are as many as updated copies
of the tuple there exist.)
With all this in place, contention related to tuples being checked by
foreign key rules should be much reduced.
As a bonus, the old behavior that a subtransaction grabbing a stronger
tuple lock than the parent (sub)transaction held on a given tuple and
later aborting caused the weaker lock to be lost, has been fixed.
Many new spec files were added for isolation tester framework, to ensure
overall behavior is sane. There's probably room for several more tests.
There were several reviewers of this patch; in particular, Noah Misch
and Andres Freund spent considerable time in it. Original idea for the
patch came from Simon Riggs, after a problem report by Joel Jacobson.
Most code is from me, with contributions from Marti Raudsepp, Alexander
Shulgin, Noah Misch and Andres Freund.
This patch was discussed in several pgsql-hackers threads; the most
important start at the following message-ids:
AANLkTimo9XVcEzfiBR-ut3KVNDkjm2Vxh+t8kAmWjPuv@mail.gmail.com1290721684-sup-3951@alvh.no-ip.org1294953201-sup-2099@alvh.no-ip.org1320343602-sup-2290@alvh.no-ip.org1339690386-sup-8927@alvh.no-ip.org4FE5FF020200002500048A3D@gw.wicourts.gov4FEAB90A0200002500048B7D@gw.wicourts.gov
Update README to explain prerequisites for
correct access to LSN fields of a page.
Independent chunk removed from checksums
patch to reduce size of patch.
Most of the replay functions for WAL record types that modify more than
one page failed to ensure that those pages were locked correctly to ensure
that concurrent queries could not see inconsistent page states. This is
a hangover from coding decisions made long before Hot Standby was added,
when it was hardly necessary to acquire buffer locks during WAL replay
at all, let alone hold them for carefully-chosen periods.
The key problem was that RestoreBkpBlocks was written to hold lock on each
page restored from a full-page image for only as long as it took to update
that page. This was guaranteed to break any WAL replay function in which
there was any update-ordering constraint between pages, because even if the
nominal order of the pages is the right one, any mixture of full-page and
non-full-page updates in the same record would result in out-of-order
updates. Moreover, it wouldn't work for situations where there's a
requirement to maintain lock on one page while updating another. Failure
to honor an update ordering constraint in this way is thought to be the
cause of bug #7648 from Daniel Farina: what seems to have happened there
is that a btree page being split was rewritten from a full-page image
before the new right sibling page was written, and because lock on the
original page was not maintained it was possible for hot standby queries to
try to traverse the page's right-link to the not-yet-existing sibling page.
To fix, get rid of RestoreBkpBlocks as such, and instead create a new
function RestoreBackupBlock that restores just one full-page image at a
time. This function can be invoked by WAL replay functions at the points
where they would otherwise perform non-full-page updates; in this way, the
physical order of page updates remains the same no matter which pages are
replaced by full-page images. We can then further adjust the logic in
individual replay functions if it is necessary to hold buffer locks
for overlapping periods. A side benefit is that we can simplify the
handling of concurrency conflict resolution by moving that code into the
record-type-specfic functions; there's no more need to contort the code
layout to keep conflict resolution in front of the RestoreBkpBlocks call.
In connection with that, standardize on zero-based numbering rather than
one-based numbering for referencing the full-page images. In HEAD, I
removed the macros XLR_BKP_BLOCK_1 through XLR_BKP_BLOCK_4. They are
still there in the header files in previous branches, but are no longer
used by the code.
In addition, fix some other bugs identified in the course of making these
changes:
spgRedoAddNode could fail to update the parent downlink at all, if the
parent tuple is in the same page as either the old or new split tuple and
we're not doing a full-page image: it would get fooled by the LSN having
been advanced already. This would result in permanent index corruption,
not just transient failure of concurrent queries.
Also, ginHeapTupleFastInsert's "merge lists" case failed to mark the old
tail page as a candidate for a full-page image; in the worst case this
could result in torn-page corruption.
heap_xlog_freeze() was inconsistent about using a cleanup lock or plain
exclusive lock: it did the former in the normal path but the latter for a
full-page image. A plain exclusive lock seems sufficient, so change to
that.
Also, remove gistRedoPageDeleteRecord(), which has been dead code since
VACUUM FULL was rewritten.
Back-patch to 9.0, where hot standby was introduced. Note however that 9.0
had a significantly different WAL-logging scheme for GIST index updates,
and it doesn't appear possible to make that scheme safe for concurrent hot
standby queries, because it can leave inconsistent states in the index even
between WAL records. Given the lack of complaints from the field, we won't
work too hard on fixing that branch.
When the "hot" members of PGPROC were split off to separate PGXACT structs,
many PGPROC fields referred to in comments were moved to PGXACT, but the
comments were neglected in the commit. Mostly this is just a search/replace
of PGPROC with PGXACT, but the way the dummy PGPROC entries are created for
prepared transactions changed more, making some of the comments totally
bogus.
Noah Misch
Removing this bit from xl_info allows us to restore the old limit of four
(not three) separate pages touched by a WAL record, which is needed for the
upcoming SP-GiST feature, and will likely be useful elsewhere in future.
When we implemented XLR_BKP_REMOVABLE in 2007, we had to do it like that
because no special WAL-visible action was taken when starting a backup.
However, now we force a segment switch when starting a backup, so a
compressing WAL archiver (such as pglesslog) that uses the state shown in
the current page header will not be fooled as to removability of backup
blocks. The only downside is that the archiver will not return to
compressing mode for up to one WAL page after the backup is over, which is
a small price to pay for getting back the extra xl_info bit. In any case
the archiver could look for XLOG_BACKUP_END records if it thought it was
worth the trouble to do so.
Bump XLOG_PAGE_MAGIC since this is effectively a change in WAL format.
Enabled by recovery_connections = on (default) and forcing archive recovery using a recovery.conf. Recovery processing now emulates the original transactions as they are replayed, providing full locking and MVCC behaviour for read only queries. Recovery must enter consistent state before connections are allowed, so there is a delay, typically short, before connections succeed. Replay of recovering transactions can conflict and in some cases deadlock with queries during recovery; these result in query cancellation after max_standby_delay seconds have expired. Infrastructure changes have minor effects on normal running, though introduce four new types of WAL record.
New test mode "make standbycheck" allows regression tests of static command behaviour on a standby server while in recovery. Typical and extreme dynamic behaviours have been checked via code inspection and manual testing. Few port specific behaviours have been utilised, though primary testing has been on Linux only so far.
This commit is the basic patch. Additional changes will follow in this release to enhance some aspects of behaviour, notably improved handling of conflicts, deadlock detection and query cancellation. Changes to VACUUM FULL are also required.
Simon Riggs, with significant and lengthy review by Heikki Linnakangas, including streamlined redesign of snapshot creation and two-phase commit.
Important contributions from Florian Pflug, Mark Kirkwood, Merlin Moncure, Greg Stark, Gianni Ciolli, Gabriele Bartolini, Hannu Krosing, Robert Haas, Tatsuo Ishii, Hiroyuki Yamada plus support and feedback from many other community members.
This patch eliminates the marking of subtransactions as SUBCOMMITTED in pg_clog
during their commit; instead they remain in-progress until main transaction
commit. At main transaction commit, the commit protocol is atomic-by-page
instead of one transaction at a time. To avoid a race condition with some
subtransactions appearing committed before others in the case where they span
more than one pg_clog page, we conserve the logic that marks them subcommitted
before marking the parent committed.
Simon Riggs with minor help from me
ReadNewTransactionId from GetSnapshotData --- with a "latestCompletedXid"
variable that is updated during transaction commit or abort. Since
latestCompletedXid is written only in places that had to lock ProcArrayLock
exclusively anyway, and is read only in places that had to lock ProcArrayLock
shared anyway, it adds no new locking requirements to the system despite being
cluster-wide. Moreover, removing ReadNewTransactionId from snapshot
acquisition eliminates the need to take both XidGenLock and ProcArrayLock at
the same time. Since XidGenLock is sometimes held across I/O this can be a
significant win. Some preliminary benchmarking suggested that this patch has
no effect on average throughput but can significantly improve the worst-case
transaction times seen in pgbench. Concept by Florian Pflug, implementation
by Tom Lane.
no need for serialization against snapshot-taking because the xact doesn't
affect anyone else's snapshot anyway. Per discussion. Also, move various
info about the interlocking of transactions and snapshots out of code comments
and into a hopefully-more-cohesive discussion in access/transam/README.
Also, remove a couple of now-obsolete comments about having to force some WAL
to be written to persuade RecordTransactionCommit to do its thing.
rows will normally never obtain an XID at all. We already did things this way
for subtransactions, but this patch extends the concept to top-level
transactions. In applications where there are lots of short read-only
transactions, this should improve performance noticeably; not so much from
removal of the actual XID-assignments, as from reduction of overhead that's
driven by the rate of XID consumption. We add a concept of a "virtual
transaction ID" so that active transactions can be uniquely identified even
if they don't have a regular XID. This is a much lighter-weight concept:
uniqueness of VXIDs is only guaranteed over the short term, and no on-disk
record is made about them.
Florian Pflug, with some editorialization by Tom.
before reporting a transaction committed. Data consistency is still
guaranteed (unlike setting fsync = off), but a crash may lose the effects
of the last few transactions. Patch by Simon, some editorialization by Tom.
misleadingly-named WriteBuffer routine, and instead require routines that
change buffer pages to call MarkBufferDirty (which does exactly what it says).
We also require that they do so before calling XLogInsert; this takes care of
the synchronization requirement documented in SyncOneBuffer. Note that
because bufmgr takes the buffer content lock (in shared mode) while writing
out any buffer, it doesn't matter whether MarkBufferDirty is executed before
the buffer content change is complete, so long as the content change is
completed before releasing exclusive lock on the buffer. So it's OK to set
the dirtybit before we fill in the LSN.
This eliminates the former kluge of needing to set the dirtybit in LockBuffer.
Aside from making the code more transparent, we can also add some new
debugging assertions, in particular that the caller of MarkBufferDirty must
hold the buffer content lock, not merely a pin.
This commit doesn't make much functional change, but it does eliminate some
duplicated code --- for instance, PageIsNew tests are now done inside
XLogReadBuffer rather than by each caller.
The GIST xlog code still needs a lot of love, but I'll worry about that
separately.
communication structure, and make it its own module with its own lock.
This should reduce contention at least a little, and it definitely makes
the code seem cleaner. Per my recent proposal.
as per recent discussions. Invent SubTransactionIds that are managed like
CommandIds (ie, counter is reset at start of each top transaction), and
use these instead of TransactionIds to keep track of subtransaction status
in those modules that need it. This means that a subtransaction does not
need an XID unless it actually inserts/modifies rows in the database.
Accordingly, don't assign it an XID nor take a lock on the XID until it
tries to do that. This saves a lot of overhead for subtransactions that
are only used for error recovery (eg plpgsql exceptions). Also, arrange
to release a subtransaction's XID lock as soon as the subtransaction
exits, in both the commit and abort cases. This avoids holding many
unique locks after a long series of subtransactions. The price is some
additional overhead in XactLockTableWait, but that seems acceptable.
Finally, restructure the state machine in xact.c to have a more orthogonal
set of states for subtransactions.
Andres Freund
Robert Haas
Alvaro Herrera
Heikki Linnakangas
Simon Riggs
Tom Lane
Peter Eisentraut
Magnus Hagander
Bruce Momjian