Replay of XLOG_BTREE_VACUUM during Hot Standby was previously thought to require
complex interlocking that matched the requirements on the master. This required
an O(N) operation that became a significant problem with large indexes, causing
replication delays of seconds or in some cases minutes while the
XLOG_BTREE_VACUUM was replayed.
This commit skips the “pin scan” that was previously required, by observing in
detail when and how it is safe to do so, with full documentation. The pin scan
is skipped only in replay; the VACUUM code path on master is not touched here.
The current commit still performs the pin scan for toast indexes, though this
can also be avoided if we recheck scans on toast indexes. Later patch will
address this.
No tests included. Manual tests using an additional patch to view WAL records
and their timing have shown the change in WAL records and their handling has
successfully reduced replication delay.
That was otherwise harmless, but tripped the new assertion in
PageGetSpecialPointer().
Reported by Amit Langote. Backpatch to 9.5, where the assertion was added.
Even though the main benefit of the Lehman and Yao algorithm for
btrees is that no locks need be held between page reads in an
index search, we were holding a buffer pin on each leaf page after
it was read until we were ready to read the next one. The reason
was so that we could treat this as a weak lock to create an
"interlock" with vacuum's deletion of heap line pointers, even
though our README file pointed out that this was not necessary for
a scan using an MVCC snapshot.
The main goal of this patch is to reduce the blocking of vacuum
processes by in-progress btree index scans (including a cursor
which is idle), but the code rearrangement also allows for one
less buffer content lock to be taken when a forward scan steps from
one page to the next, which results in a small but consistent
performance improvement in many workloads.
This patch leaves behavior unchanged for some cases, which can be
addressed separately so that each case can be evaluated on its own
merits. These unchanged cases are when a scan uses a non-MVCC
snapshot, an index-only scan, and a scan of a btree index for which
modifications are not WAL-logged. If later patches allow all of
these cases to drop the buffer pin after reading a leaf page, then
the btree vacuum process can be simplified; it will no longer need
the "super-exclusive" lock to delete tuples from a page.
Reviewed by Heikki Linnakangas and Kyotaro Horiguchi
xlog.c is huge, this makes it a little bit smaller, which is nice. Functions
related to putting together the WAL record are in xloginsert.c, and the
lower level stuff for managing WAL buffers and such are in xlog.c.
Also move the definition of XLogRecord to a separate header file. This
causes churn in the #includes of all the files that write WAL records, and
redo routines, but it avoids pulling in xlog.h into most places.
Reviewed by Michael Paquier, Alvaro Herrera, Andres Freund and Amit Kapila.
log_newpage is used by many indexams, in addition to heap, but for
historical reasons it's always been part of the heapam rmgr. Starting with
9.3, we have another WAL record type for logging an image of a page,
XLOG_FPI. Simplify things by moving log_newpage and log_newpage_buffer to
xlog.c, and switch to using the XLOG_FPI record type.
Bump the WAL version number because the code to replay the old HEAP_NEWPAGE
records is removed.
The previous code, perhaps out of concern for avoid memory leaks, formed
the tuple in one memory context and then copied it to another memory
context. However, this doesn't appear to be necessary, since
index_form_tuple and the functions it calls take precautions against
leaking memory. In my testing, building the tuple directly inside the
sort context shaves several percent off the index build time.
Rearrange things so we do that.
Patch by me. Review by Amit Kapila, Tom Lane, Andres Freund.
In short, we don't allow a page to be deleted if it's the rightmost child
of its parent, but that situation can change after we check for it.
Problem
-------
We check that the page to be deleted is not the rightmost child of its
parent, and then lock its left sibling, the page itself, its right sibling,
and the parent, in that order. However, if the parent page is split after
the check but before acquiring the locks, the target page might become the
rightmost child, if the split happens at the right place. That leads to an
error in vacuum (I reproduced this by setting a breakpoint in debugger):
ERROR: failed to delete rightmost child 41 of block 3 in index "foo_pkey"
We currently re-check that the page is still the rightmost child, and throw
the above error if it's not. We could easily just give up rather than throw
an error, but that approach doesn't scale to half-dead pages. To recap,
although we don't normally allow deleting the rightmost child, if the page
is the *only* child of its parent, we delete the child page and mark the
parent page as half-dead in one atomic operation. But before we do that, we
check that the parent can later be deleted, by checking that it in turn is
not the rightmost child of the grandparent (potentially recursing all the
way up to the root). But the same situation can arise there - the
grandparent can be split while we're not holding the locks. We end up with
a half-dead page that we cannot delete.
To make things worse, the keyspace of the deleted page has already been
transferred to its right sibling. As the README points out, the keyspace at
the grandparent level is "out-of-whack" until the half-dead page is deleted,
and if enough tuples with keys in the transferred keyspace are inserted, the
page might get split and a downlink might be inserted into the grandparent
that is out-of-order. That might not cause any serious problem if it's
transient (as the README ponders), but is surely bad if it stays that way.
Solution
--------
This patch changes the page deletion algorithm to avoid that problem. After
checking that the topmost page in the chain of to-be-deleted pages is not
the rightmost child of its parent, and then deleting the pages from bottom
up, unlink the pages from top to bottom. This way, the intermediate stages
are similar to the intermediate stages in page splitting, and there is no
transient stage where the keyspace is "out-of-whack". The topmost page in
the to-be-deleted chain doesn't have a downlink pointing to it, like a page
split before the downlink has been inserted.
This also allows us to get rid of the cleanup step after WAL recovery, if we
crash during page deletion. The deletion will be continued at next VACUUM,
but the tree is consistent for searches and insertions at every step.
This bug is old, all supported versions are affected, but this patch is too
big to back-patch (and changes the WAL record formats of related records).
We have not heard any reports of the bug from users, so clearly it's not
easy to bump into. Maybe backpatch later, after this has had some field
testing.
Reviewed by Kevin Grittner and Peter Geoghegan.
These flushes were added in my commit d2896a9ed, which added the btree
logic that keeps a cached copy of the index metapage data in index relcache
entries. The idea was to ensure that other backends would promptly update
their cached copies after a change. However, this is not really necessary,
since _bt_getroot() has adequate defenses against believing a stale root
page link, and _bt_getrootheight() doesn't have to be 100% right.
Moreover, if it were necessary, a relcache flush would be an unreliable way
to do it, since the sinval mechanism believes that relcache flush requests
represent transactional updates, and therefore discards them on transaction
rollback. Therefore, we might as well drop these flush requests and save
the time to rebuild the whole relcache entry after a metapage change.
If we ever try to support in-place truncation of btree indexes, it might
be necessary to revisit this issue so that _bt_getroot() can't get caught
by trying to follow a metapage link to a page that no longer exists.
A possible solution to that is to make use of an smgr, rather than
relcache, inval request to force other backends to discard their cached
metapages. But for the moment this is not worth pursuing.
In ordinary operation, VACUUM must be careful to take a cleanup lock on
each leaf page of a btree index; this ensures that no indexscans could
still be "in flight" to heap tuples due to be deleted. (Because of
possible index-tuple motion due to concurrent page splits, it's not enough
to lock only the pages we're deleting index tuples from.) In Hot Standby,
the WAL replay process must likewise lock every leaf page. There were
several bugs in the code for that:
* The replay scan might come across unused, all-zero pages in the index.
While btree_xlog_vacuum itself did the right thing (ie, nothing) with
such pages, xlogutils.c supposed that such pages must be corrupt and
would throw an error. This accounts for various reports of replication
failures with "PANIC: WAL contains references to invalid pages". To
fix, add a ReadBufferMode value that instructs XLogReadBufferExtended
not to complain when we're doing this.
* btree_xlog_vacuum performed the extra locking if standbyState ==
STANDBY_SNAPSHOT_READY, but that's not the correct test: we won't open up
for hot standby queries until the database has reached consistency, and
we don't want to do the extra locking till then either, for fear of reading
corrupted pages (which bufmgr.c would complain about). Fix by exporting a
new function from xlog.c that will report whether we're actually in hot
standby replay mode.
* To ensure full coverage of the index in the replay scan, btvacuumscan
would emit a dummy WAL record for the last page of the index, if no
vacuuming work had been done on that page. However, if the last page
of the index is all-zero, that would result in corruption of said page,
since the functions called on it weren't prepared to handle that case.
There's no need to lock any such pages, so change the logic to target
the last normal leaf page instead.
The first two of these bugs were diagnosed by Andres Freund, the other one
by me. Fixes based on ideas from Heikki Linnakangas and myself.
This has been wrong since Hot Standby was introduced, so back-patch to 9.0.
MarkBufferDirtyHint() writes WAL, and should know if it's got a
standard buffer or not. Currently, the only callers where buffer_std
is false are related to the FSM.
In passing, rename XLOG_HINT to XLOG_FPI, which is more descriptive.
Back-patch to 9.3.
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.
This patch addresses the problem that applications currently have to
extract object names from possibly-localized textual error messages,
if they want to know for example which index caused a UNIQUE_VIOLATION
failure. It adds new error message fields to the wire protocol, which
can carry the name of a table, table column, data type, or constraint
associated with the error. (Since the protocol spec has always instructed
clients to ignore unrecognized field types, this should not create any
compatibility problem.)
Support for providing these new fields has been added to just a limited set
of error reports (mainly, those in the "integrity constraint violation"
SQLSTATE class), but we will doubtless add them to more calls in future.
Pavel Stehule, reviewed and extensively revised by Peter Geoghegan, with
additional hacking by Tom Lane.
This fixes another error in commit 9e8da0f757.
I neglected to make the mark/restore functionality save and restore the
current set of array key values, which led to strange behavior if an
IndexScan with ScalarArrayOpExpr quals was used as the inner side of a
mergejoin. Per bug #7570 from Melese Tesfaye.
The heapam XLog functions are used by other modules, not all of which
are interested in the rest of the heapam API. With this, we let them
get just the XLog stuff in which they are interested and not pollute
them with unrelated includes.
Also, since heapam.h no longer requires xlog.h, many files that do
include heapam.h no longer get xlog.h automatically, including a few
headers. This is useful because heapam.h is getting pulled in by
execnodes.h, which is in turn included by a lot of files.
The need for this was debated when we put in the index-only-scan feature,
but at the time we had no near-term expectation of having AMs that could
support such scans for only some indexes; so we kept it simple. However,
the SP-GiST AM forces the issue, so let's fix it.
This patch only installs the new API; no behavior actually changes.
In general the data returned by an index-only scan should have the
datatypes originally computed by FormIndexDatum. If the index opclasses
use "storage" datatypes different from their input datatypes, the scan
tuple will not have the same rowtype attributed to the index; but we had
a hard-wired assumption that that was true in nodeIndexonlyscan.c. We'd
already hacked around the issue for the one case where the types are
different in btree indexes (btree name_ops), but this would definitely
come back to bite us if we ever implement index-only scans in GiST.
To fix, require the index AM to explicitly provide the tupdesc for the
tuple it is returning. btree can just pass back the index's tupdesc, but
GiST will have to work harder when and if it supports index-only scans.
I had previously proposed fixing this by allowing the index AM to fill the
scan tuple slot directly; but on reflection that seemed like a module
layering violation, since TupleTableSlots are creatures of the executor.
At least in the btree case, it would also be less efficient, since the
tuple deconstruction work would occur even for rows later found to be
invisible to the scan's snapshot.
It's been bothering me for several days that pretending that the cstring
data stored in a btree name_ops column is really a "name" Datum could lead
to reading past the end of memory. However, given the current memory
layout used for index-only scans in the btree code, a crash is in fact not
possible. Document that so we don't break it. I have not thought of any
other solutions that aren't fairly ugly too, and most of them lose the
functionality of index-only scans on name columns altogether, so this seems
like the way to go.
We copy all the matched tuples off the page during _bt_readpage, instead of
expensively re-locking the page during each subsequent tuple fetch. This
costs a bit more local storage, but not more than 2*BLCKSZ worth, and the
reduction in LWLock traffic is certainly worth that. What's more, this
lets us get rid of the API wart in the original patch that said an index AM
could randomly decline to supply an index tuple despite having asserted
pg_am.amcanreturn. That will be important for future improvements in the
index-only-scan feature, since the executor will now be able to rely on
having the index data available.
When a btree index contains all columns required by the query, and the
visibility map shows that all tuples on a target heap page are
visible-to-all, we don't need to fetch that heap page. This patch depends
on the previous patches that made the visibility map reliable.
There's a fair amount left to do here, notably trying to figure out a less
chintzy way of estimating the cost of an index-only scan, but the core
functionality seems ready to commit.
Robert Haas and Ibrar Ahmed, with some previous work by Heikki Linnakangas.
Change the way UPDATEs are handled. Instead of maintaining a chain of
tuple-level locks in shared memory, copy any existing locks on the old
tuple to the new tuple at UPDATE. Any existing page-level lock needs to
be duplicated too, as a lock on the new tuple. That was neglected
previously.
Store xmin on tuple-level predicate locks, to distinguish a lock on an old
already-recycled tuple from a new tuple at the same physical location.
Failure to distinguish them caused loops in the tuple-lock chains, as
reported by YAMAMOTO Takashi. Although we don't use the chain representation
of UPDATEs anymore, it seems like a good idea to store the xmin to avoid
some false positives if no other reason.
CheckSingleTargetForConflictsIn now correctly handles the case where a lock
that's being held is not reflected in the local lock table. That happens
if another backend acquires a lock on our behalf due to an UPDATE or a page
split.
PredicateLockPageCombine now retains locks for the page that is being
removed, rather than removing them. This prevents a potentially dangerous
false-positive inconsistency where the local lock table believes that a lock
is held, but it is actually not.
Dan Ports and Kevin Grittner
Until now, our Serializable mode has in fact been what's called Snapshot
Isolation, which allows some anomalies that could not occur in any
serialized ordering of the transactions. This patch fixes that using a
method called Serializable Snapshot Isolation, based on research papers by
Michael J. Cahill (see README-SSI for full references). In Serializable
Snapshot Isolation, transactions run like they do in Snapshot Isolation,
but a predicate lock manager observes the reads and writes performed and
aborts transactions if it detects that an anomaly might occur. This method
produces some false positives, ie. it sometimes aborts transactions even
though there is no anomaly.
To track reads we implement predicate locking, see storage/lmgr/predicate.c.
Whenever a tuple is read, a predicate lock is acquired on the tuple. Shared
memory is finite, so when a transaction takes many tuple-level locks on a
page, the locks are promoted to a single page-level lock, and further to a
single relation level lock if necessary. To lock key values with no matching
tuple, a sequential scan always takes a relation-level lock, and an index
scan acquires a page-level lock that covers the search key, whether or not
there are any matching keys at the moment.
A predicate lock doesn't conflict with any regular locks or with another
predicate locks in the normal sense. They're only used by the predicate lock
manager to detect the danger of anomalies. Only serializable transactions
participate in predicate locking, so there should be no extra overhead for
for other transactions.
Predicate locks can't be released at commit, but must be remembered until
all the transactions that overlapped with it have completed. That means that
we need to remember an unbounded amount of predicate locks, so we apply a
lossy but conservative method of tracking locks for committed transactions.
If we run short of shared memory, we overflow to a new "pg_serial" SLRU
pool.
We don't currently allow Serializable transactions in Hot Standby mode.
That would be hard, because even read-only transactions can cause anomalies
that wouldn't otherwise occur.
Serializable isolation mode now means the new fully serializable level.
Repeatable Read gives you the old Snapshot Isolation level that we have
always had.
Kevin Grittner and Dan Ports, reviewed by Jeff Davis, Heikki Linnakangas and
Anssi Kääriäinen
The contents of an unlogged table are WAL-logged; thus, they are not
available on standby servers and are truncated whenever the database
system enters recovery. Indexes on unlogged tables are also unlogged.
Unlogged GiST indexes are not currently supported.
This is a heavily revised version of builtin_knngist_core-0.9. The
ordering operators are no longer mixed in with actual quals, which would
have confused not only humans but significant parts of the planner.
Instead, ordering operators are carried separately throughout planning and
execution.
Since the API for ambeginscan and amrescan functions had to be changed
anyway, this commit takes the opportunity to rationalize that a bit.
RelationGetIndexScan no longer forces a premature index_rescan call;
instead, callers of index_beginscan must call index_rescan too. Aside from
making the AM-side initialization logic a bit less peculiar, this has the
advantage that we do not make a useless extra am_rescan call when there are
runtime key values. AMs formerly could not assume that the key values
passed to amrescan were actually valid; now they can.
Teodor Sigaev and Tom Lane
WAL record for btree delete contains a list of tids, even when backup
blocks are present. We follow the tids to their heap tuples, taking
care to follow LP_REDIRECT tuples. We ignore LP_DEAD tuples on the
understanding that they will always have xmin/xmax earlier than any
LP_NORMAL tuples referred to by killed index tuples. Iff all tuples
are LP_DEAD we return InvalidTransactionId. The heap relfilenode is
added to the WAL record, requiring API changes to pass down the heap
Relation. XLOG_PAGE_MAGIC updated.
VACUUM FULL INPLACE), along with a boatload of subsidiary code and complexity.
Per discussion, the use case for this method of vacuuming is no longer large
enough to justify maintaining it; not to mention that we don't wish to invest
the work that would be needed to make it play nicely with Hot Standby.
Aside from the code directly related to old-style VACUUM FULL, this commit
removes support for certain WAL record types that could only be generated
within VACUUM FULL, redirect-pointer removal in heap_page_prune, and
nontransactional generation of cache invalidation sinval messages (the last
being the sticking point for Hot Standby).
We still have to retain all code that copes with finding HEAP_MOVED_OFF and
HEAP_MOVED_IN flag bits on existing tuples. This can't be removed as long
as we want to support in-place update from pre-9.0 databases.
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.
The current implementation fires an AFTER ROW trigger for each tuple that
looks like it might be non-unique according to the index contents at the
time of insertion. This works well as long as there aren't many conflicts,
but won't scale to massive unique-key reassignments. Improving that case
is a TODO item.
Dean Rasheed
behavior in cases where we don't know the heap tuple count accurately; in
particular partial vacuum, but this also makes the API a bit more useful
for ANALYZE. This patch adds "estimated_count" flags to both structs so
that an approximate count can be flagged as such, and adjusts the logic
so that approximate counts are not used for updating pg_class.reltuples.
This fixes my previous complaint that VACUUM was putting ridiculous values
into pg_class.reltuples for indexes. The actual impact of that bug is
limited, because the planner only pays attention to reltuples for an index
if the index is partial; which probably explains why beta testers hadn't
noticed a degradation in plan quality from it. But it needs to be fixed.
The whole thing is a bit messy and should be redesigned in future, because
reltuples now has the potential to drift quite far away from reality when
a long period elapses with no non-partial vacuums. But this is as good as
it's going to get for 8.4.
Simon Riggs
Bruce Momjian
Heikki Linnakangas
Andres Freund
Kevin Grittner
Robert Haas
Tom Lane
Jeff Davis
Alvaro Herrera
Magnus Hagander