* Preserve task result in TaskHandle by keeping join handle around
The solution is not great, but it should hep to debug staging issue
I tried to do it in a least destructive way. TaskHandle used only in
one place so it is ok to use something less generic unless we want
to extend its usage across the codebase. In its current current form
for its single usage place it looks too abstract
Some problems around this code:
1. Task can drop event sender and continue running
2. Task cannot be joined several times (probably not needed,
but still, can be surprising)
3. Had to split task event into two types because ahyhow::Error
does not implement clone. So TaskContinueEvent derives clone
but usual task evend does not. Clone requirement appears
because we clone the current value in next_task_event.
Taking it by reference is complicated.
4. Split between Init and Started is artificial and comes from
watch::channel requirement to have some initial value.
To summarize from 3 and 4. It may be a better idea to use
RWLock or a bounded channel instead
Instead of spawning helper threads, we now use Tokio tasks. There
are multiple Tokio runtimes, for different kinds of tasks. One for
serving libpq client connections, another for background operations
like GC and compaction, and so on. That's not strictly required, we
could use just one runtime, but with this you can still get an
overview of what's happening with "top -H".
There's one subtle behavior in how TenantState is updated. Before this
patch, if you deleted all timelines from a tenant, its GC and
compaction loops were stopped, and the tenant went back to Idle
state. We no longer do that. The empty tenant stays Active. The
changes to test_tenant_tasks.py are related to that.
There's still plenty of synchronous code and blocking. For example, we
still use blocking std::io functions for all file I/O, and the
communication with WAL redo processes is still uses low-level unix
poll(). We might want to rewrite those later, but this will do for
now. The model is that local file I/O is considered to be fast enough
that blocking - and preventing other tasks running in the same thread -
is acceptable.
Move all the fields that were returned by the wal_receiver endpoint into
timeline_detail. Internally, move those fields from the separate global
WAL_RECEIVERS hash into the LayeredTimeline struct. That way, all the
information about a timeline is kept in one place.
In the passing, I noted that the 'thread_id' field was removed from
WalReceiverEntry in commit e5cb727572, but it forgot to update
openapi_spec.yml. This commit removes that too.
What the WAL receiver really connects to is the safekeeper. The
"producer" term is a bit misleading, as the safekeeper doesn't produce
the WAL, the compute node does.
This change also applies to the name of the field used in the mgmt API
in in the response of the
'/v1/tenant/:tenant_id/timeline/:timeline_id/wal_receiver' endpoint.
AFAICS that's not used anywhere else than one python test, so it
should be OK to change it.
download operations of all timelines for one tenant are now grouped
together so when attach is invoked pageserver downloads all of them
and registers them in a single apply_sync_status_update call so
branches can be used safely with attach/detach
I noticed that the pageserver has a very large virtual memory size,
several GB, even though it doesn't actually use that much
memory. That's not much of a problem normally, but I hit it because I
wanted to run tests with a limited virtual memory size, by calling
setrlimit(RLIMIT_AS), but the highest limit you can set is 2 GB. I was
not able to start pageserver with a limit of 2 GB.
On Linux, each thread allocates 32 MB of virtual memory. I read this
on some random forum on the Internet, but unfortunately could not find
the source again now. Empirically, reducing the number of threads clearly
helps to bring down the virtual memory size.
Aside from the virtual memory usage, it seems excessive to launch 40
threads in both of those thread pools. The tokio default is to have as
many worker threads as there are CPU cores in the system. That seems
like a fine heuristic for us, too, so remove the explicit setting of
the pool size and rely on the default. Note that the GC and compaction
tasks are actually run with tokio spawn_blocking, so the threads that
are actually doing the work, and possibly waiting on I/O, are not
consuming threads from the thread pool. The WAL receiver work is done
in the tokio worker threads, but the WAL receivers are more CPU bound
so that seems OK.
Also remove the explicit maxinum on blocking tasks. I'm not sure what
the right value for that would be, or whether the value we set (100)
would be better than the tokio default (512). Since the value was
arbitrary, let's just rely on the tokio default for that, too.
* Avoid reconnecting to safekeeper immediately after its failure by limiting candidates to those with fewest connection attempts. Thus we don't have to wait lagging_wal_timeout (10s by default) before switch happens even if no new changes are generated, and current test_restarts_under_load expects some commits to happen within 4s.
* Make default max_lsn_wal_lag larger, otherwise we constant reconnections happen during normal work.
* Fix wal_connection_attempts maintanance, preventing busy loop of reconnections.
Resolves#1488.
- implemented `GET tenant/:tenant_id/timeline/:timeline_id/wal_receiver` endpoint
- returned `thread_id` in `thread_mgr::spawn`
- added `latest_gc_cutoff_lsn` field to `LocalTimelineInfo` struct
Add tenant config API and 'zenith tenant config' CLI command.
Add 'show' query to pageserver protocol for tenantspecific config parameters
Refactoring: move tenant_config code to a separate module.
Save tenant conf file to tenant's directory, when tenant is created to recover it on pageserver restart.
Ignore error during tenant config loading, while it is not supported by console
Define PiTR interval for GC.
refer #1320
This is a backwards-incompatible change. The new pageserver cannot
read repositories created with an old pageserver binary, or vice
versa.
Simplify Repository to a value-store
------------------------------------
Move the responsibility of tracking relation metadata, like which
relations exist and what are their sizes, from Repository to a new
module, pgdatadir_mapping.rs. The interface to Repository is now a
simple key-value PUT/GET operations.
It's still not any old key-value store though. A Repository is still
responsible from handling branching, and every GET operation comes
with an LSN.
Mapping from Postgres data directory to keys/values
---------------------------------------------------
All the data is now stored in the key-value store. The
'pgdatadir_mapping.rs' module handles mapping from PostgreSQL objects
like relation pages and SLRUs, to key-value pairs.
The key to the Repository key-value store is a Key struct, which
consists of a few integer fields. It's wide enough to store a full
RelFileNode, fork and block number, and to distinguish those from
metadata keys.
'pgdatadir_mapping.rs' is also responsible for maintaining a
"partitioning" of the keyspace. Partitioning means splitting the
keyspace so that each partition holds a roughly equal number of keys.
The partitioning is used when new image layer files are created, so
that each image layer file is roughly the same size.
The partitioning is also responsible for reclaiming space used by
deleted keys. The Repository implementation doesn't have any explicit
support for deleting keys. Instead, the deleted keys are simply
omitted from the partitioning, and when a new image layer is created,
the omitted keys are not copied over to the new image layer. We might
want to implement tombstone keys in the future, to reclaim space
faster, but this will work for now.
Changes to low-level layer file code
------------------------------------
The concept of a "segment" is gone. Each layer file can now store an
arbitrary range of Keys.
Checkpointing, compaction
-------------------------
The background tasks are somewhat different now. Whenever
checkpoint_distance is reached, the WAL receiver thread "freezes" the
current in-memory layer, and creates a new one. This is a quick
operation and doesn't perform any I/O yet. It then launches a
background "layer flushing thread" to write the frozen layer to disk,
as a new L0 delta layer. This mechanism takes care of durability. It
replaces the checkpointing thread.
Compaction is a new background operation that takes a bunch of L0
delta layers, and reshuffles the data in them. It runs in a separate
compaction thread.
Deployment
----------
This also contains changes to the ansible scripts that enable having
multiple different pageservers running at the same time in the staging
environment. We will use that to keep an old version of the pageserver
running, for clusters created with the old version, at the same time
with a new pageserver with the new binary.
Author: Heikki Linnakangas
Author: Konstantin Knizhnik <knizhnik@zenith.tech>
Author: Andrey Taranik <andrey@zenith.tech>
Reviewed-by: Matthias Van De Meent <matthias@zenith.tech>
Reviewed-by: Bojan Serafimov <bojan@zenith.tech>
Reviewed-by: Konstantin Knizhnik <knizhnik@zenith.tech>
Reviewed-by: Anton Shyrabokau <antons@zenith.tech>
Reviewed-by: Dhammika Pathirana <dham@zenith.tech>
Reviewed-by: Kirill Bulatov <kirill@zenith.tech>
Reviewed-by: Anastasia Lubennikova <anastasia@zenith.tech>
Reviewed-by: Alexey Kondratov <alexey@zenith.tech>
Use log::error!() instead. I spotted a few of these "connection error"
lines in the logs, without timestamps and the other stuff we print for
all other log messages.
to pass current_timeline_size to compute node
Put standby_status_update fields into ZenithFeedback and send them as one message.
Pass values sizes together with keys in ZenithFeedback message.
This patch includes attach/detach http endpoints in pageservers. Some
changes in callmemaybe handling inside safekeeper and an integrational
test to check migration with and without load. There are still some
rough edges that will be addressed in follow up patches
This introduces a new module to handle thread creation and shutdown.
All page server threads are now registered in a global hash map, and
there's a function to request individual threads to shut down gracefully.
Thread shutdown request is signalled to the thread with a flag, as well
as a Future that can be used to wake up async operations if shutdown is
requested. Use that facility to have the libpq listener thread respond
to pageserver shutdown, based on Kirill's earlier prototype
(https://github.com/zenithdb/zenith/pull/1088). That addresses
https://github.com/zenithdb/zenith/issues/1036, previously the libpq
listener thread would not exit until one more connection arrives.
This also eliminates a resource leak in the accept() loop. Previously,
we added the JoinHanlde of each new thread to a vector but old handles
for threads that had already exited were never removed.
A timeline ID is only guaranteed to be unique for a particular tenant,
so you need to use tenant ID + timeline ID as the key, rather than just
timeline ID.
The safekeeper currently makes the same assumption, and we should fix that
too, but this commit just addresses this one case in the page server.
In the passing, reorder some function arguments to be more consistent.
This patch allows to shutdown wal receiver when there are no messages
and wal receiver is blocked inside tokio-postgres. In this case it
cannot check the shutdown flag.
This patch switches to use async interface of tokio-postgres directly
without sync wrappers. It opens the possibility to use tokio::select!
between the phsycal_stream.next() and a shutdown channel readiness to
interrupt replication process.
Also this allows to shutdown only particular wal receiver without
using global shutdown_requested flag.
Change meaning of lsns in HOT_STANDBY_FEEDBACK:
flush_lsn = disk_consistent_lsn,
apply_lsn = remote_consistent_lsn
Update compute node backpressure configuration respectively.
Update compute node configuration:
set 'synchronous_commit=remote_write' in setup without safekeepers.
This way compute node doesn't have to wait for data checkpoint on pageserver.
This doesn't guarantee data durability, but we only use this setup for tests, so it's fine.
Rename save_decoded_record() to ingest_record(), and move the
responsibility for decoding the record into ingest_record().
Also move the responsibility of updating the CheckPoint relish to
ingest_record(). Put it in a new WalIngest struct, to help with tracking
that.
Move the code for decoding a WAL stream into WAL records into
'postgres_ffi', and keep the code to parse the WAL records deeper in
'pageserver' crate, renamed to walrecord.rs.
This tidies up the dependencies a bit. 'walkeeper' reuses the same
waldecoder routines, and it used to depend on 'pageserver' because of
that. Now it only depends on 'postgres_ffi'.
(The comment in walkeeper/Cargo.toml that claimed that the dependency was
needed for ZTimelineId was obsolete. ZTimelineId is defined in
'zenith_utils', the dependency was actually needed for the waldecoder.)
write_lsn - The last LSN received and processed by pageserver's walreceiver.
flush_lsn - same as write_lsn. At pageserver it doesn't guarantees data persistence, but it's fine. We rely on safekeepers.
apply_lsn - The LSN at which pageserver guaranteed persistence of all received data (disk_consistent_lsn).
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.
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>
