Part of #8130
## Problem
Pageserver previously goes through the kernel page cache for all the
IOs. The kernel page cache makes light-loaded pageserver have deceptive
fast performance. Using direct IO would offer predictable latencies of
our virtual file IO operations.
In particular for reads, the data pages also have an extremely low
temporal locality because the most frequently accessed pages are cached
on the compute side.
## Summary of changes
This PR enables pageserver to use direct IO for delta layer and image
layer reads. We can ship them separately because these layers are
write-once, read-many, so we will not be mixing buffered IO with direct
IO.
- implement `IoBufferMut`, an buffer type with aligned allocation
(currently set to 512).
- use `IoBufferMut` at all places we are doing reads on image + delta
layers.
- leverage Rust type system and use `IoBufAlignedMut` marker trait to
guarantee that the input buffers for the IO operations are aligned.
- page cache allocation is also made aligned.
_* in-memory layer reads and the write path will be shipped separately._
## Testing
Integration test suite run with O_DIRECT enabled:
https://github.com/neondatabase/neon/pull/9350
## Performance
We evaluated performance based on the `get-page-at-latest-lsn`
benchmark. The results demonstrate a decrease in the number of IOps, no
sigificant change in the latency mean, and an slight improvement on the
p99.9 and p99.99 latencies.
[Benchmark](https://www.notion.so/neondatabase/Benchmark-O_DIRECT-for-image-and-delta-layers-2024-10-01-112f189e00478092a195ea5a0137e706?pvs=4)
## Rollout
We will add `virtual_file_io_mode=direct` region by region to enable
direct IO on image + delta layers.
Signed-off-by: Yuchen Liang <yuchen@neon.tech>
part of Epic https://github.com/neondatabase/neon/issues/7386
# Motivation
The materialized page cache adds complexity to the code base, which
increases the maintenance burden and risk for subtle and hard to
reproduce bugs such as #8050.
Further, the best hit rate that we currently achieve in production is ca
1% of materialized page cache lookups for
`task_kind=PageRequestHandler`. Other task kinds have hit rates <0.2%.
Last, caching page images in Pageserver rewards under-sized caches in
Computes because reading from Pageserver's materialized page cache over
the network is often sufficiently fast (low hundreds of microseconds).
Such Computes should upscale their local caches to fit their working
set, rather than repeatedly requesting the same page from Pageserver.
Some more discussion and context in internal thread
https://neondb.slack.com/archives/C033RQ5SPDH/p1718714037708459
# Changes
This PR removes the materialized page cache code & metrics.
The infrastructure for different key kinds in `PageCache` is left in
place, even though the "Immutable" key kind is the only remaining one.
This can be further simplified in a future commit.
Some tests started failing because their total runtime was dependent on
high materialized page cache hit rates. This test makes them
fixed-runtime or raises pytest timeouts:
* test_local_file_cache_unlink
* test_physical_replication
* test_pg_regress
# Performance
I focussed on ensuring that this PR will not result in a performance
regression in prod.
* **getpage** requests: our production metrics have shown the
materialized page cache to be irrelevant (low hit rate). Also,
Pageserver is the wrong place to cache page images, it should happen in
compute.
* **ingest** (`task_kind=WalReceiverConnectionHandler`): prod metrics
show 0 percent hit rate, so, removing will not be a regression.
* **get_lsn_by_timestamp**: important API for branch creation, used by
control pane. The clog pages that this code uses are not
materialize-page-cached because they're not 8k. No risk of introducing a
regression here.
We will watch the various nightly benchmarks closely for more results
before shipping to prod.
Nightly has added a bunch of compiler and linter warnings. There is also
two dependencies that fail compilation on latest nightly due to using
the old `stdsimd` feature name. This PR fixes them.
## Problem
In https://github.com/neondatabase/neon/pull/5957, the most essential
types were updated to use TenantShardId rather than TenantId. That
unblocked other work, but didn't fully enable running multiple shards
from the same tenant on the same pageserver.
## Summary of changes
- Use TenantShardId in page cache key for materialized pages
- Update mgr.rs get_tenant() and list_tenants() functions to use a shard
id, and update all callers.
- Eliminate the exactly_one_or_none helper in mgr.rs and all code that
used it
- Convert timeline HTTP routes to use tenant_shard_id
Note on page cache:
```
struct MaterializedPageHashKey {
/// Why is this TenantShardId rather than TenantId?
///
/// Usually, the materialized value of a page@lsn is identical on any shard in the same tenant. However, this
/// this not the case for certain internally-generated pages (e.g. relation sizes). In future, we may make this
/// key smaller by omitting the shard, if we ensure that reads to such pages always skip the cache, or are
/// special-cased in some other way.
tenant_shard_id: TenantShardId,
timeline_id: TimelineId,
key: Key,
}
```
These changes help with identifying thrashing.
The existing `pageserver_page_cache_find_victim_iters_total` is already
useful, but, it doesn't tell us how many individual find_victim() calls
are happening, only how many clock-LRU steps happened in the entire
system,
without info about whether we needed to actually evict other data vs
just scan for a long time, e.g., because the cache is large.
The changes in this PR allows us to
1. count each possible outcome separately, esp evictions
2. compute mean iterations/outcome
I don't think anyone except me was paying close attention to
`pageserver_page_cache_find_victim_iters_total` before, so,
I think the slight behavior change of also counting iterations
for the 'iters exceeded' case is fine.
refs https://github.com/neondatabase/cloud/issues/8351
refs https://github.com/neondatabase/neon/issues/5479
Problem
=======
Prior to this PR, when we had a cache miss, we'd get back a write guard,
fill it, the drop it and retry the read from cache.
If there's severe contention for the cache, it could happen that the
just-filled data gets evicted before our retry, resulting in lost work
and no forward progress.
Solution
========
This PR leverages the now-available `tokio::sync::RwLockWriteGuard`'s
`downgrade()` functionality to turn the filled slot write guard into a
read guard.
We don't drop the guard at any point, so, forward progress is ensured.
Refs
====
Stacked atop https://github.com/neondatabase/neon/pull/5480
part of https://github.com/neondatabase/neon/issues/4743
specifically part of https://github.com/neondatabase/neon/issues/5479
It is wasteful to cycle through the page cache slots trying to find a
victim slot if all the slots are currently un-evictable because a read /
write guard is alive.
We suspect this wasteful cycling to be the root cause for an
"indigestion" we observed in staging (#5291).
The hypothesis is that we `.await` after we get ahold of a read / write
guard, and that tokio actually deschedules us in favor of another
future.
If that other future then needs a page slot, it can't get ours because
we're holding the guard.
Repeat this, and eventually, the other future(s) will find themselves
doing `find_victim` until they hit `exceeded evict iter limit`.
The `find_victim` is wasteful and CPU-starves the futures that are
already holding the read/write guard. A `yield` inside `find_victim`
could mitigate the starvation, but wouldn't fix the wasting of CPU
cycles.
So instead, this PR queues waiters behind a tokio semaphore that counts
evictable slots.
The downside is that this stops the clock page replacement if we have 0
evictable slots.
Also, as explained by the big block comment in `find_victims`, the
semaphore doesn't fully prevent starvation because because we can't make
tokio prioritize those tasks executing `find_victim` that have been
trying the longest.
Implementation
===============
We need to acquire the semaphore permit before locking the slot.
Otherwise, we could deadlock / discover that all permits are gone and
would have to relinquish the slot, having moved forward the Clock LRU
without making progress.
The downside is that, we never get full throughput for read-heavy
workloads, because, until the reader coalesces onto an existing permit,
it'll hold its own permit.
Addendum To Root-Cause Analysis In #5291
========================================
Since merging that PR, @arpad-m pointed out that we couldn't have
reached the `slot.write().await` with his patches because the
VirtualFile slots can't have all been write-locked, because we only hold
them locked while the IO is ongoing, and the IO is still done with
synchronous system calls in that patch set, so, we can have had at most
$number_of_executor_threads locked at any given time.
I count 3 tokio runtimes that do `Timeline::get`, each with 8 executor
threads in our deployment => $number_of_executor_threads = 3*8 = 24 .
But the virtual file cache has 100 slots.
We both agree that nothing changed about the core hypothesis, i.e.,
additional await points inside VirtualFile caused higher concurrency
resulting in exhaustion of page cache slots.
But we'll need to reproduce the issue and investigate further to truly
understand the root cause, or find out that & why we were indeed using
100 VirtualFile slots.
TODO: could it be compaction that needs to hold guards of many
VirtualFile's in its iterators?
The TaskKind dimension added in #5339 is insufficient to understand what
kind of data causes the cache hits.
Regarding performance considerations: I'm not too worried because we're
moving from 3 to 4 one-byte sized fields; likely the space now used by
the new field was padding before. Didn't check this, though, and it
doesn't matter, we need the data.
What I don't like about this PR is that we have an `Unknown` content
type, and I also don't like that there's no compile-time way to assert
that it's set to something != `Unknown` when calling the page cache.
But, this is what I could come up with before tomorrow’s release, and I
think it covers the hot paths.
This PR adds a `task_kind` label to page cache access metrics.
These are to validate our hypothesis that the high hit page cache rate
we observe in prod is due to internal tasks, not getpage requests from
compute.
We believe the latter should near-always be a pageserver-page-cache
_miss_ because compute has it's own page cache, and hence there is no
locality of reference for its accesses to pageserver page cache.
Before this PR, we didn't have `RequestContext` propagation to any code
below the on-demand downloader.
The vast majority of changes in this PR is concerned with adding that
propagation.
It was easy to interpret comment in the page cache initialization code
to be about justifying why we leak here at all, not just why this
specific type of leaking is done (which the comment was actually meant
to describe).
See
https://github.com/neondatabase/neon/pull/5125#discussion_r1308445993
---------
Co-authored-by: Joonas Koivunen <joonas@neon.tech>
## Problem
`read_blk` does I/O and thus we would like to make it async. We can't
make the function async as long as the `PageReadGuard` returned by
`read_blk` isn't `Send`. The page cache is called by `read_blk`, and
thus it can't be async without `read_blk` being async. Thus, we have a
circular dependency.
## Summary of changes
Due to the circular dependency, we convert both the page cache and
`read_blk` to async at the same time:
We make the page cache use `tokio::sync` synchronization primitives as
those are `Send`. This makes all the places that acquire a lock require
async though, which we then also do. This includes also asyncification
of the `read_blk` function.
Builds upon #4994, #5015, #5056, and #5129.
Part of #4743.
## Problem
I saw these things while working on #5111.
## Summary of changes
* Add a comment explaining why we use `Vec::leak` instead of
`Vec::into_boxed_slice` plus `Box::leak`.
* Add another comment explaining what `valid` is doing, it wasn't very
clear before.
* Add a function `set_usage_count` to not set it directly.
Before this patch, when dropping an EphemeralFile, we'd scan the entire
`slots` to proactively evict its pages (`drop_buffers_for_immutable`).
This was _necessary_ before #4994 because the page cache was a
write-back cache: we'd be deleting the EphemeralFile from disk after,
so, if we hadn't evicted its pages before that, write-back in
`find_victim` wouldhave failed.
But, since #4994, the page cache is a read-only cache, so, it's safe
to keep read-only data cached. It's never going to get accessed again
and eventually, `find_victim` will evict it.
The only remaining advantage of `drop_buffers_for_immutable` over
relying on `find_victim` is that `find_victim` has to do the clock
page replacement iterations until the count reaches 0,
whereas `drop_buffers_for_immutable` can kick the page out right away.
However, weigh that against the cost of `drop_buffers_for_immutable`,
which currently scans the entire `slots` array to find the
EphemeralFile's pages.
Alternatives have been proposed in #5122 and #5128, but, they come
with their own overheads & trade-offs.
Also, the real reason why we're looking into this piece of code is
that we want to make the slots rwlock async in #5023.
Since `drop_buffers_for_immutable` is called from drop, and there
is no async drop, it would be nice to not have to deal with this.
So, let's just stop doing `drop_buffers_for_immutable` and observe
the performance impact in benchmarks.
(This PR is the successor of https://github.com/neondatabase/neon/pull/4984 )
## Summary
The current way in which `EphemeralFile` uses `PageCache` complicates
the Pageserver code base to a degree that isn't worth it.
This PR refactors how we cache `EphemeralFile` contents, by exploiting
the append-only nature of `EphemeralFile`.
The result is that `PageCache` only holds `ImmutableFilePage` and
`MaterializedPage`.
These types of pages are read-only and evictable without write-back.
This allows us to remove the writeback code from `PageCache`, also
eliminating an entire failure mode.
Futher, many great open-source libraries exist to solve the problem of a
read-only cache,
much better than our `page_cache.rs` (e.g., better replacement policy,
less global locking).
With this PR, we can now explore using them.
## Problem & Analysis
Before this PR, `PageCache` had three types of pages:
* `ImmutableFilePage`: caches Delta / Image layer file contents
* `MaterializedPage`: caches results of Timeline::get (page
materialization)
* `EphemeralPage`: caches `EphemeralFile` contents
`EphemeralPage` is quite different from `ImmutableFilePage` and
`MaterializedPage`:
* Immutable and materialized pages are for the acceleration of (future)
reads of the same data using `PAGE_CACHE_SIZE * PAGE_SIZE` bytes of
DRAM.
* Ephemeral pages are a write-back cache of `EphemeralFile` contents,
i.e., if there is pressure in the page cache, we spill `EphemeralFile`
contents to disk.
`EphemeralFile` is only used by `InMemoryLayer`, for the following
purposes:
* **write**: when filling up the `InMemoryLayer`, via `impl BlobWriter
for EphemeralFile`
* **read**: when doing **page reconstruction** for a page@lsn that isn't
written to disk
* **read**: when writing L0 layer files, we re-read the `InMemoryLayer`
and put the contents into the L0 delta writer
(**`create_delta_layer`**). This happens every 10min or when
InMemoryLayer reaches 256MB in size.
The access patterns of the `InMemoryLayer` use case are as follows:
* **write**: via `BlobWriter`, strictly append-only
* **read for page reconstruction**: via `BlobReader`, random
* **read for `create_delta_layer`**: via `BlobReader`, dependent on
data, but generally random. Why?
* in classical LSM terms, this function is what writes the
memory-resident `C0` tree into the disk-resident `C1` tree
* in our system, though, the values of InMemoryLayer are stored in an
EphemeralFile, and hence they are not guaranteed to be memory-resident
* the function reads `Value`s in `Key, LSN` order, which is `!=` insert
order
What do these `EphemeralFile`-level access patterns mean for the page
cache?
* **write**:
* the common case is that `Value` is a WAL record, and if it isn't a
full-page-image WAL record, then it's smaller than `PAGE_SIZE`
* So, the `EphemeralPage` pages act as a buffer for these `< PAGE_CACHE`
sized writes.
* If there's no page cache eviction between subsequent
`InMemoryLayer::put_value` calls, the `EphemeralPage` is still resident,
so the page cache avoids doing a `write` system call.
* In practice, a busy page server will have page cache evictions because
we only configure 64MB of page cache size.
* **reads for page reconstruction**: read acceleration, just as for the
other page types.
* **reads for `create_delta_layer`**:
* The `Value` reads happen through a `BlockCursor`, which optimizes the
case of repeated reads from the same page.
* So, the best case is that subsequent values are located on the same
page; hence `BlockCursor`s buffer is maximally effective.
* The worst case is that each `Value` is on a different page; hence the
`BlockCursor`'s 1-page-sized buffer is ineffective.
* The best case translates into `256MB/PAGE_SIZE` page cache accesses,
one per page.
* the worst case translates into `#Values` page cache accesses
* again, the page cache accesses must be assumed to be random because
the `Value`s aren't accessed in insertion order but `Key, LSN` order.
## Summary of changes
Preliminaries for this PR were:
- #5003
- #5004
- #5005
- uncommitted microbenchmark in #5011
Based on the observations outlined above, this PR makes the following
changes:
* Rip out `EphemeralPage` from `page_cache.rs`
* Move the `block_io::FileId` to `page_cache::FileId`
* Add a `PAGE_SIZE`d buffer to the `EphemeralPage` struct.
It's called `mutable_tail`.
* Change `write_blob` to use `mutable_tail` for the write buffering
instead of a page cache page.
* if `mutable_tail` is full, it writes it out to disk, zeroes it out,
and re-uses it.
* There is explicitly no double-buffering, so that memory allocation per
`EphemeralFile` instance is fixed.
* Change `read_blob` to return different `BlockLease` variants depending
on `blknum`
* for the `blknum` that corresponds to the `mutable_tail`, return a ref
to it
* Rust borrowing rules prevent `write_blob` calls while refs are
outstanding.
* for all non-tail blocks, return a page-cached `ImmutablePage`
* It is safe to page-cache these as ImmutablePage because EphemeralFile
is append-only.
## Performance
How doe the changes above affect performance?
M claim is: not significantly.
* **write path**:
* before this PR, the `EphemeralFile::write_blob` didn't issue its own
`write` system calls.
* If there were enough free pages, it didn't issue *any* `write` system
calls.
* If it had to evict other `EphemeralPage`s to get pages a page for its
writes (`get_buf_for_write`), the page cache code would implicitly issue
the writeback of victim pages as needed.
* With this PR, `EphemeralFile::write_blob` *always* issues *all* of its
*own* `write` system calls.
* Also, the writes are explicit instead of implicit through page cache
write back, which will help #4743
* The perf impact of always doing the writes is the CPU overhead and
syscall latency.
* Before this PR, we might have never issued them if there were enough
free pages.
* We don't issue `fsync` and can expect the writes to only hit the
kernel page cache.
* There is also an advantage in issuing the writes directly: the perf
impact is paid by the tenant that caused the writes, instead of whatever
tenant evicts the `EphemeralPage`.
* **reads for page reconstruction**: no impact.
* The `write_blob` function pre-warms the page cache when it writes the
`mutable_tail` to disk.
* So, the behavior is the same as with the EphemeralPages before this
PR.
* **reads for `create_delta_layer`**: no impact.
* Same argument as for page reconstruction.
* Note for the future:
* going through the page cache likely causes read amplification here.
Why?
* Due to the `Key,Lsn`-ordered access pattern, we don't read all the
values in the page before moving to the next page. In the worst case, we
might read the same page multiple times to read different `Values` from
it.
* So, it might be better to bypass the page cache here.
* Idea drafts:
* bypass PS page cache + prefetch pipeline + iovec-based IO
* bypass PS page cache + use `copy_file_range` to copy from ephemeral
file into the L0 delta file, without going through user space
This makes it more explicit that these are different u64-sized
namespaces.
Re-using one in place of the other would be catastrophic.
Prep for https://github.com/neondatabase/neon/pull/4994
which will eliminate the ephemeral_file::FileId and move the
blob_io::FileId into page_cache.
It makes sense to have this preliminary commit though,
to minimize amount of new concept in #4994 and other
preliminaries that depend on that work.
Re-export only things that are used by other modules.
In the future, I'm imagining that we run bindgen twice, for Postgres
v14 and v15. The two sets of bindings would go into separate
'bindings_v14' and 'bindings_v15' modules.
Rearrange postgres_ffi modules.
Move function, to avoid Postgres version dependency in timelines.rs
Move function to generate a logical-message WAL record to postgres_ffi.
This introduces two new abstraction layers for I/O:
- Block I/O, and
- Blob I/O.
The BlockReader trait abstracts a file or something else that can be read
in 8kB pages. It is implemented by EphemeralFiles, and by a new
FileBlockReader struct that allows reading arbitrary VirtualFiles in that
manner, utilizing the page cache.
There is also a new BlockCursor struct that works as a cursor over a
BlockReader. When you create a BlockCursor and read the first page using
it, it keeps the reference to the page. If you access the same page again,
it avoids going to page cache and quickly returns the same page again.
That can save a lot of lookups in the page cache if you perform multiple
reads.
The Blob-oriented API allows reading and writing "blobs" of arbitrary
length. It is a layer on top of the block-oriented API. When you write
a blob with the write_blob() function, it writes a length field
followed by the actual data to the underlying block storage, and
returns the offset where the blob was stored. The blob can be
retrieved later using the offset.
Finally, this replaces the I/O code in image-, delta-, and in-memory
layers to use the new abstractions. These replace the 'bookfile'
crate.
This is a backwards-incompatible change to the storage format.
It happened in unit tests. If a thread tries to read a buffer while
already holding a lock on one buffer, the code to find a victim buffer
to evict could try to evict the buffer that's already locked. To fix,
skip locked buffers.
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>
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.
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>
Once upon a time, 'page_cache.rs' contained an actual page cache, but
it hasn't for a very long time. Rename to reflect what it actually does
these days.
Now that we only have one Repository implementation, no need for the
command-line options to choose it either. I'm removing these as a separate
commit to show what we will need to do if we add another Repository
implementation in the future (even though I don't foresee us doing that
any time soon)
The layered storage format is good enough that we don't need the rocksdb
implementation anymore. There are a lot of known issues but we'll keep
working on them.
This replaces the RocksDB based implementation with an approach using
"snapshot files" on disk, and in-memory btreemaps to hold the recent
changes.
This make the repository implementation a configuration option. You can
choose 'layered' or 'rocksdb' with "zenith init --repository-format=<format>"
The unit tests have been refactored to exercise both implementations.
'layered' is now the default.
Push/pull is not implemented. The 'test_history_inmemory' test has been
commented out accordingly. It's not clear how we will implement that
functionality; probably by copying the snapshot files directly.
Current state with authentication.
Page server validates JWT token passed as a password during connection
phase and later when performing an action such as create branch tenant
parameter of an operation is validated to match one submitted in token.
To allow access from console there is dedicated scope: PageServerApi,
this scope allows access to all tenants. See code for access validation in:
PageServerHandler::check_permission.
Because we are in progress of refactoring of communication layer
involving wal proposer protocol, and safekeeper<->pageserver. Safekeeper
now doesn’t check token passed from compute, and uses “hardcoded” token
passed via environment variable to communicate with pageserver.
Compute postgres now takes token from environment variable and passes it
as a password field in pageserver connection. It is not passed through
settings because then user will be able to retrieve it using pg_settings
or SHOW ..
I’ve added basic test in test_auth.py. Probably after we add
authentication to remaining network paths we should enable it by default
and switch all existing tests to use it.
The codepath for tenant_create command first launched the WAL redo
thread, and then called branches::create_repo() which checked if the
tenant's directory already exists. That's problematic, because
launching the WAL redo thread will run initdb if the directory doesn't
already exist. Race condition: If the tenant already exists, it will
have a WAL redo thread already running, and the old and new WAL redo
thread might try to run initdb at the same time, causing all kinds of
weird failures.
The test_pageserver_api test was failing 100% repeatably on my laptop
because of this. I'm not sure why this doesn't occur on the CI:
Jul 31 18:05:48.877 INFO running initdb in "./tenants/5227e4eb90894775ac6b8a8c76f24b2e/wal-redo-datadir", location: pageserver::walredo, pageserver/src/walredo.rs:483
thread 'WAL redo thread' panicked at 'initdb failed: The files belonging to this database system will be owned by user "heikki".
This user must also own the server process.
The database cluster will be initialized with locale "C".
The default database encoding has accordingly been set to "SQL_ASCII".
The default text search configuration will be set to "english".
Data page checksums are disabled.
creating directory ./tenants/0305b1326f3ea33add0929d516da7cb6/wal-redo-datadir ... ok
creating subdirectories ... ok
selecting dynamic shared memory implementation ... posix
selecting default max_connections ... 100
selecting default shared_buffers ... 128MB
selecting default time zone ... Europe/Helsinki
creating configuration files ... ok
running bootstrap script ...
stderr:
2021-07-31 15:05:48.875 GMT [282569] LOG: could not open configuration file "/home/heikki/git-sandbox/zenith/test_output/test_tenant_list/repo/./tenants/0305b1326f3ea33add0929d516da7cb6/wal-redo-datadir/postgresql.conf": No such file or directory
2021-07-31 15:05:48.875 GMT [282569] FATAL: configuration file "/home/heikki/git-sandbox/zenith/test_output/test_tenant_list/repo/./tenants/0305b1326f3ea33add0929d516da7cb6/wal-redo-datadir/postgresql.conf" contains errors
child process exited with exit code 1
initdb: removing data directory "./tenants/0305b1326f3ea33add0929d516da7cb6/wal-redo-datadir"
this patch adds support for tenants. This touches mostly pageserver.
Directory layout on disk is changed to contain new layer of indirection.
Now path to particular repository has the following structure: <pageserver workdir>/tenants/<tenant
id>. Tenant id has the same format as timeline id. Tenant id is included in
pageserver commands when needed. Also new commands are available in
pageserver: tenant_list, tenant_create. This is also reflected CLI.
During init default tenant is created and it's id is saved in CLI config,
so following commands can use it without extra options. Tenant id is also included in
compute postgres configuration, so it can be passed via ServerInfo to
safekeeper and in connection string to pageserver.
For more info see docs/multitenancy.md.
- All timelines are now stored in the same rocksdb repository. The GET
functions have been taught to follow the ancestors.
- Change the way relation size is stored. Instead of inserting "tombstone"
entries for blocks that are truncated away, store relation size as
separate key-value entry for each relation
- Add an abstraction for the key-value store: ObjectStore. It allows
swapping RocksDB with some other key-value store easily. Perhaps we
will write our own storage implementation using that interface, or
perhaps we'll need a different abstraction, but this is a small
improvement over status quo in any case.
- Garbage Collection is broken and commented out. It's not clear where and
how it should be implemented.
It's created once early in server startup, after parsing the
command-line options, and never modified afterwards. To simplify
things, pass it around as static ref, instead of making copies in all
the different structs. We still pass around a reference to it, rather
than putting it in a global variable, to allow unit testing with
different configs in the same process.
This patch started as an effort to support CLI working against remote
pageserver, but turned into a pretty big refactoring.
* CLI now does not look into repository files directly. New commands
'branch_create' and 'identify_system' were introduced into page_service to
support that.
* Branch management that was scattered between local_env and
zenith/main.rs is moved into pageserver/branches.rs. That code could better fit
in Repository/Timeline impl, but I'll leave that for a different patch.
* All tests-related code from local_env went into integration_tests/src/lib.rs as an
extension to PostgresNode trait.
* Paths-generating functions were concentrated around corresponding config
types (LocalEnv and PageserverConf).
Turn WalRedoManager into an abstract trait, so that it can be easily
mocked in unit tests.
One change here is that the WAL redo manager is no longer tied to a
specific zenith timeline. It didn't do anything with that information
aside from using it in the dummy datadir's name. We could use any
random string for that purpose, it's just to prevent two WAL redo
managers from stepping over each other. But this commit actually
changes things so that all timelines use the same WAL redo manager, so
that's not necessary. We will probably want to maintain a pool of WAL
redo processes in the future, but for now let's keep it simple.
In the passing, fix some comments.
We used to create them under .zenith/.zenith/<timelineid>. The double
.zenith was clearly not intentional. Change it to
.zenith/timelines/<timelineid>.
Fixes https://github.com/zenithdb/zenith/issues/127
This moves things around:
- The PageCache is split into two structs: Repository and Timeline. A
Repository holds multiple Timelines. In order to get a page version,
you must first get a reference to the Repository, then the Timeline
in the repository, and finally call the get_page_at_lsn() function
on the Timeline object. This sounds complicated, but because each
connection from a compute node, and each WAL receiver, only deals
with one timeline at a time, the callers can get the reference to
the Timeline object once and hold onto it. The Timeline corresponds
most closely to the old PageCache object.
- Repository and Timeline are now abstract traits, so that we can
support multiple implementations. I don't actually expect us to have
multiple implementations for long. We have the RocksDB
implementation now, but as soon as we have a different
implementation that's usable, I expect that we will retire the
RocksDB implementation. But I think this abstraction works as good
documentation in any case: it's now easier to see what the interface
for storing and loading pages from the repository is, by looking at
the Repository/Timeline traits. They abstract traits are in
repository.rs, and the RocksDB implementation of them is in
repository/rocksdb.rs.
- page_cache.rs is now a "switchboard" to get a handle to the
repository. Currently, the page server can only handle one
repository at a time, so there isn't much there, but in the future
we might do multi-tenancy there.
Make the caller of request_redo() responsible for gathering the WAL records
to redo, and for storing the reconstructed page image back in the page
cache. This leaves the WAL redo manager purely responsible for dealing with
the postgres child process, removing its dependency on the PageCache.
