This doesnt show up at the moment, because we never create a delta
layer with end-LSN equal to the last LSN. We always create an image
layer at that LSN instead. For example, if the latest processed LSN is
100, we would create a delta layer with end LSN 100 (exclusive), and
an image layer at 100. But that's just how InMemoryLayer::write_to_disk
happens to work at the moment, there's no fundamental reason it needs
to always create that image layer. I noticed this bug when I tried to
change the logic in InMemoryLayer::write_to_disk to only create an
image layer after a few delta layers.
The "in-memory layer" is misnomer now, each in-memory layer is now actually
backed by a file. The files are ephemeral, in that they don't survive page
server crash or shutdown.
To avoid reading the file for every operation,
"ephemeral files" are cached in a page cache.
This includes changes from 'inmemory-layer-chunks' branch to serialize /
the page versions when they are added to the open layer. The difference is
that they are not serialized to the expandable in-memory "chunk buffer", but
written out to the file.
Out of scope LSNs include pre initdb LSNs, and LSNs prior to
latest_gc_cutoff.
To get there there was also two cleanups:
* Fix error handling in Execute message handler. This fixes behaviour
when basebackup retured an error. Previously pageserver thread just
died.
* Remove "ancestor" file which previously contained ancestor id and
branch lsn. Currently the same data can be obtained from metadata file.
And just the way we handled ancestor file in the code introduced the
case when branching fails timeline directory is created but there is no data in it
except ancestor file. And this confused gc because it scans
directories. So it is better to just remove ancestor file and clean up
this timeline directory creation so it happens after all validity
checks have passed
Ever since we've had frozen in-memory layers, having an 'end_lsn' no
longer means that the layer has been dropped. Need to check the 'dropped'
flag explicitly.
This was reliably causing a failure on the new 'test_parallel_copy' test
in https://github.com/zenithdb/zenith/pull/864. I'm not sure why it
doesn't happen on main branch, but the bug is pretty straightforward when
you see it.
This introduces new timeline field latest_gc_cutoff. It is updated
before each gc iteration. New check is added to branch_timelines to
prevent branch creation with start point less than latest_gc_cutoff.
Also this adds a check to get_page_at_lsn which asserts that lsn at
which the page is requested was not garbage collected. This check
currently is triggered for readonly nodes which are pinned to specific
lsn and because they are not tracked in pageserver garbage collection
can remove data that still might be referenced. This is a bug and will
be fixed separately.
The buffer cache is shared across all tenants, allowing memory to be
dynamically allocated where it's needed the most. The cache works on 8 kB
pages, and uses the clock algorithm for replacement policy; same as the
PostgreSQL buffer cache.
One peculiarity is that the materialized page versions can be looked up
by an inexact LSN, to find the latest page version with an LSN >= the
search key.
The code is structured to support caching other kinds of pages in the same
cache in the future, but with a different mapping key.
Co-authored-by: Patrick Insinger <patrick@zenith.tech>
During parallel load of a table, Postgres sometimes requests a page from
the page server for which no WAL has been generated yet. That's normal;
Postgres expects the page to be full of zeros. There was a special case
for that in LayeredTimeline::materialize_page, but the problem remained
when you're crossing a segment boundary, so that there's no layer for
the segment at all.
It would be nice to have a more robust cross-check for this case. That
might need help from the Postgres side. But this extends the bandaid fix
we had in materialize_page() to the case where cross segment boundary.
Fixes https://github.com/zenithdb/zenith/issues/841
There were two separate locking issues that could lead to a deadlock,
both related to holding a lock for longer than necessary:
1. In the loop in `VirtualFile::with_file`, the "handle_guard" was
held across iterations of the loop. Because of that, if the handle was
changed by a concurrent thread, the loop would try to acquire the
handle lock, when it was still holding the lock from previous
iteration. To fix, release the lock earlier. There was no need to hold
it across iterations, it was just accidental.
2. In the same function, we also held the "slot_guard" longer than
necessary. It's only needed in the first part of the loop, where we
check if the current handle is valid. If it's not, the slot lock can
be immediately released. But it was not, it was kept over the
acquisition of the handle lock. I'm not sure if that alone could cause
problems, but let's release the lock as soon as possible anyway.
Add a test case, based on Konstantin's test program to demonstrate the
deadlock.
Currently, whenever a page version is needed from an image or delta
layer, we open the file and read and parse the bookfile headers. That's
pretty expensive. To reduce the overhead, introduce a cache of open file
descriptors, and use that to cache the Book objects so that we don't need
to read the metadata on every access.
Commit 960c7d69a8 changed the LSN returned in the Continue case in
InMemoryLayer::get_page_reconstruct_data(), but neglected to make the
same change in DeltaLayer.
Also add an escape hatch to the loop in materialize_page() to avoid
getting stuck in an infinite loop, if a bug like this reoccurs.
Git commit sha is displayed when --version flag is used and is written
to logs during service startup. Uses git_version crate when git is
available, and GIT_VERSION environment variable otherwise which is the case for docker
builds.
The tokio futures added some overhead, so switch to plain non-blocking
I/O with poll(). In a simple pgbench test on my laptop (select-only
queries, scale-factor 1 `pgbench -P1 -T50 -S`), this gives about 10%
improvement, from about 4300 TPS to 4800 TPS.
Instead of building a separate Vec<u8> to hold each message, serialize all
the messages to one big Vec<u8>. This eliminates some Vec allocation and
memcpy() overhead. The downside is that if there are a lot of records to
replay, we have to serialize them all into one big chunk of memory.
That shouldn't be a problem in practice. If you need to replay millions
of records to reconstruct a page, we should've materialized a new image
of that page earlier already.
We might want to have custom serialize/deserialize functions for
WALRecords and PageVersions for performance reasons, see github issue 832.
But that would probably look a bit different from this, and currently
these functions are just dead.
Adds simple global tracking of memory used by the in-memory layers. It's
very approximate, it doesn't take into account allocator, memory
fragmentation or many other things, but it's a good first step.
After storing a WAL record in the repository, the WAL receiver checks
if the global memory usage. If it's above a configurable threshold (hard
coded at 128 MB at the moment), it evicts a layer. The victim layer is
chosen by GClock algorithm, similar to that used in the Postgres buffer
cache.
This stops the page server from using an unbounded amount of memory. It's
pretty crude, the eviction and materializing and writing a layer to disk
happens now in the WAL receiver thread. It would be nice to move that
to a background thread, and it would be nice to have a smarter policy on
when to materialize a new image layer and when to just write out a delta
layer, and it would be nice to have more accurate accounting of memory.
But this should fix the most pressing OOM issues, and is a step in the
right direction.
Co-authored-by: Patrick Insinger <patrickinsinger@gmail.com>
This calculation is not that heavy but it is needed only in tests, and
in case the number of tenants/timelines is high the calculation can take
noticeable time.
Resolves https://github.com/zenithdb/zenith/issues/804
Change 'zenith.signal' file to a human-readable format, similar to
backup_label. It can contain a "PREV LSN: %X/%X" line, or a special
value to indicate that it's OK to start with invalid LSN ('none'), or
that it's a read-only node and generating WAL is forbidden
('invalid').
The 'zenith pg create' and 'zenith pg start' commands now take a node
name parameter, separate from the branch name. If the node name is not
given, it defaults to the branch name, so this doesn't break existing
scripts.
If you pass "foo@<lsn>" as the branch name, a read-only node anchored
at that LSN is created. The anchoring is performed by setting the
'recovery_target_lsn' option in the postgresql.conf file, and putting
the server into standby mode with 'standby.signal'.
We no longer store the synthetic checkpoint record in the WAL segment.
The postgres startup code has been changed to use the copy of the
checkpoint record in the pg_control file, when starting in zenith
mode.
Previously, the first WAL record on the 'main' branch overwrote the
initial checkpoint record, with invalid 'xl_prev'. That's harmless, but
also pretty ugly. I bumped into this while I was trying to tighen up the
checks for when a valid 'prev_lsn' is required. With this patch, the
first WAL record gets a valid 'xl_prev' value. It doesn't matter much
currently, but let's be tidy.
