## Problem
There are some common types that we pass into tenants and timelines as
we construct them, such as remote storage and the broker client.
Currently the list is small, but this is likely to grow -- the deletion
queue PR (#4960) pushed some methods to the point of clippy complaining
they had too many args, because of the extra deletion queue client being
passed around.
There are some shared objects that currently aren't passed around
explicitly because they use a static `once_cell` (e.g.
CONCURRENT_COMPACTIONS), but as we add more resource management and
concurreny control over time, it will be more readable & testable to
pass a type around in the respective Resources object, rather than to
coordinate via static objects. The `Resources` structures in this PR
will make it easier to add references to central coordination functions,
without having to rely on statics.
## Summary of changes
- For `Tenant`, the `broker_client` and `remote_storage` are bundled
into `TenantSharedResources`
- For `Timeline`, the `remote_client` is wrapped into
`TimelineResources`.
Both of these structures will get an additional deletion queue member in
#4960.
## Problem
IndexPart contains two redundant lists of layer names: a set of the
names, and then a map of name to metadata.
We already required that all the layers in `timeline_layers` are also in
`layers_metadata`, in `initialize_with_current_remote_index_part`, so if
there were any index_part.json files in the field that relied on these
sets being different, they would already be broken.
## Summary of changes
`timeline_layers` is made private and no longer read at runtime. It is
still serialized, but not deserialized.
`disk_consistent_lsn` is also made private, as this field only exists
for convenience of humans reading the serialized JSON.
This prepares us to entirely remove `timeline_layers` in a future
release, once this change is fully deployed, and therefore no
pageservers are trying to read the field.
I'm still a bit nervous about attach -> crash case. But it should work.
(unlike case with timeline). Ideally would be cool to cover this with
test.
This continues tradition of adding bool flags for Tenant::set_stopping.
Probably lifecycle project will help with fixing it.
## Problem
Before, DeltaLayer dumping (via `cargo run --release -p pagectl --
print-layer-file` ) would crash as one can't call `Handle::block_on` in
an async executor thread.
## Summary of changes
Avoid the problem by using `DeltaLayerInner::load_keys` to load the keys
into RAM (which we already do during compaction), and then load the
values one by one during dumping.
- move them to pageserver which is the only dependant on the crate fail
- "move" the exported macro to the new module
- support at init time the same failpoints as runtime
Found while debugging test failures and making tests more repeatable by
allowing "exit" from pageserver start via environment variables. Made
those changes to `test_gc_cutoff.py`.
---------
Co-authored-by: Christian Schwarz <christian@neon.tech>
(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
## Problem
The performance benchmark in `test_runner/performance/test_layer_map.py`
is currently failing due to the warning added in #4888.
## Summary of changes
The test mentioned has a `compaction_target_size` of 8192, which is just
one page size. This is an unattainable goal, as we generate at least
three pages: one for the header, one for the b-tree (minimally sized
ones have just the root node in a single page), one for the data.
Therefore, we add two pages to the warning limit. The warning text
becomes a bit less accurate but I think this is okay.
I will have to change these as I change remote_timeline_client api in
#4938. So a bit of cleanup, handle my comments which were just resolved
during initial review.
Cleanup:
- use unwrap in tests instead of mixed `?` and `unwrap`
- use `Handle` instead of `&'static Reactor` to make the
RemoteTimelineClient more natural
- use arrays in tests
- use plain `#[tokio::test]`
Rather temporary solution before proper:
https://github.com/neondatabase/neon/issues/5006
It requires more plumbing so lets not attach deleted tenants first and
then implement resume.
Additionally fix `assert_prefix_empty`. It had a buggy prefix calculation,
and since we always asserted for absence of stuff it worked. Here I
started to assert for presence of stuff too and it failed. Added more
"presence" asserts to other places to be confident that it works.
Resolves [#5016](https://github.com/neondatabase/neon/issues/5016)
Before this patch, we had the `off` and `blknum` as function-wide
mutable state. Now it's contained in the `Writer` struct.
The use of `push_bytes` instead of index-based filling of the buffer
also makes it easier to reason about what's going on.
This is prep for https://github.com/neondatabase/neon/pull/4994
Restores #4937 work relating to the ability to use `ResidentDeltaLayer`
(which is an Arc wrapper) in #4938 for the ValueRef's by removing the
borrow from `ValueRef` and providing it from an upper layer.
This should not have any functional changes, most importantly, the
`main` will continue to use the borrowed `DeltaLayerInner`. It might be
that I can change #4938 to be like this. If that is so, I'll gladly rip
out the `Ref` and move the borrow back. But I'll first want to look at
the current test failures.
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.
## Problem
PR #4839 has already reduced the number of b-tree traversals and vec
creations from 3 to 2, but as pointed out in
https://github.com/neondatabase/neon/pull/4839#discussion_r1279167815 ,
we would ideally just traverse the b-tree once during compaction.
Afer #4836, the two vecs created are one for the list of keys, lsns and
sizes, and one for the list of `(key, lsn, value reference)`. However,
they are not equal, as pointed out in
https://github.com/neondatabase/neon/pull/4839#issuecomment-1660418012
and the following comment: the key vec creation combines multiple
entries for which the lsn is changing but the key stays the same into
one, with the size being the sum of the sub-sizes. In SQL, this would
correspond to something like `SELECT key, lsn, SUM(size) FROM b_tree
GROUP BY key;` and `SELECT key, lsn, val_ref FROM b_tree;`. Therefore,
the join operation is non-trivial.
## Summary of changes
This PR merges the two lists of keys and value references into one. It's
not a trivial change and affects the size pattern of the resulting
files, which is why this is in a separate PR from #4839 .
The key vec is used in compaction for determining when to start a new
layer file. The loop uses various thresholds to come to this conclusion,
but the grouping via the key has led to the behaviour that regardless of
the threshold, it only starts a new file when either a new key is
encountered, or a new delta file.
The new code now does the combination after the merging and sorting of
the various keys from the delta files. This *mostly* does the same as
the old code, except for a detail: with the grouping done on a
per-delta-layer basis, the sorted and merged vec would still have
multiple entries for multiple delta files, but now, we don't have an
easy way to tell when a new input delta layer file is encountered, so we
cannot create multiple entries on that basis easily.
To prevent possibly infinite growth, our new grouping code compares the
combined size with the threshold, and if it is exceeded, it cuts a new
entry so that the downstream code can cut a new output file. Here, we
perform a tradeoff however, as if the threshold is too small, we risk
putting entries for the same key into multiple layer files, but if the
threshold is too big, we can in some instances exceed the target size.
Currently, we set the threshold to the target size, so in theory we
would stay below or roughly at double the `target_file_size`.
We also fix the way the size was calculated for the last key. The calculation
was wrong and accounted for the old layer's btree, even though we
already account for the overhead of the in-construction btree.
Builds on top of #4839 .
Previously list_prefixes was incorrectly used for that purpose. Change
to use list_files. Add a test.
Some drive by refactorings on python side to move helpers out of
specific test file to be widely accessible
resolves https://github.com/neondatabase/neon/issues/4499
## Problem
Deletions can be possibly reordered. Use fsync to avoid the case when
mark file doesnt exist but other tenant/timeline files do.
See added comments.
resolves#4987
## Problem
The `BlockReader` trait is not ready to be asyncified, as associated
types are not supported by asyncification strategies like via the
`async_trait` macro, or via adopting enums.
## Summary of changes
Remove the `BlockLease` associated type from the `BlockReader` trait and
turn it into an enum instead, bearing the same name. The enum has two
variants, one of which is gated by `#[cfg(test)]`. Therefore, outside of
test settings, the enum has zero overhead over just having the
`PageReadGuard`. Using the enum allows us to impl `BlockReader` without
needing the page cache.
Part of https://github.com/neondatabase/neon/issues/4743
## Problem
The `BlockCursor::read_blob` and `BlockCursor::read_blob_into_buf`
functions are calling `read_blk` internally, so if we want to make that
function async fn, they need to be async themselves.
## Summary of changes
* We first turn `ValueRef::load` into an async fn.
* Then, we switch the `RwLock` implementation in `InMemoryLayer` to use
the one from `tokio`.
* Last, we convert the `read_blob` and `read_blob_into_buf` functions
into async fn.
In three instances we use `Handle::block_on`:
* one use is in compaction code, which currently isn't async. We put the
entire loop into an `async` block to prevent the potentially hot loop
from doing cross-thread operations.
* one use is in dumping code for `DeltaLayer`. The "proper" way to
address this would be to enable the visit function to take async
closures, but then we'd need to be generic over async fs non async,
which [isn't supported by rust right
now](https://blog.rust-lang.org/inside-rust/2022/07/27/keyword-generics.html).
The other alternative would be to do a first pass where we cache the
data into memory, and only then to dump it.
* the third use is in writing code, inside a loop that copies from one
file to another. It is is synchronous and we'd like to keep it that way
(for now?).
Part of #4743
Originated from test failure where we got SlowDown error from s3.
The patch generalizes `download_retry` to not be download specific.
Resulting `retry` function is moved to utils crate. `download_retries`
is now a thin wrapper around this `retry` function.
To ensure that all needed retries are in place test code now uses
`test_remote_failures=1` setting.
Ref https://neondb.slack.com/archives/C059ZC138NR/p1691743624353009
## Problem
Currently, image generation reads delta layers before writing out
subsequent image layers, which updates the access time of the delta
layers and effectively puts them at the back of the queue for eviction.
This is the opposite of what we want, because after a delta layer is
covered by a later image layer, it's likely that subsequent reads of
latest data will hit the image rather than the delta layer, so the delta
layer should be quite a good candidate for eviction.
## Summary of changes
`RequestContext` gets a new `ATimeBehavior` field, and a
`RequestContextBuilder` helper so that we can optionally add the new
field without growing `RequestContext::new` every time we add something
like this.
Request context is passed into the `record_access` function, and the
access time is not updated if `ATimeBehavior::Skip` is set.
The compaction background task constructs its request context with this
skip policy.
Closes: https://github.com/neondatabase/neon/issues/4969
## Problem
In some places, the lock on `InMemoryLayerInner` is only created to
obtain `end_lsn`. This is not needed however, if we move `end_lsn` to
`InMemoryLayer` instead.
## Summary of changes
Make `end_lsn` a member of `InMemoryLayer`, and do less locking of
`InMemoryLayerInner`. `end_lsn` is changed from `Option<Lsn>` into an
`OnceLock<Lsn>`. Thanks to this change, we don't need to lock any more
in three functions.
Part of #4743 . Suggested in
https://github.com/neondatabase/neon/pull/4905#issuecomment-1666458428 .
## Problem
The `DiskBtreeReader::visit` function calls `read_blk` internally, and
while #4863 converted the API of `visit` to async, the internal function
is still recursive. So, analogously to #4838, we turn the recursive
function into an iterative one.
## Summary of changes
First, we prepare the change by moving the for loop outside of the case
switch, so that we only have one loop that calls recursion. Then, we
switch from using recursion to an approach where we store the search
path inside the tree on a stack on the heap.
The caller of the `visit` function can control when the search over the
B-Tree ends, by returning `false` from the closure. This is often used
to either only find one specific entry (by always returning `false`),
but it is also used to iterate over all entries of the B-tree (by always
returning `true`), or to look for ranges (mostly in tests, but
`get_value_reconstruct_data` also has such a use).
Each stack entry contains two things: the block number (aka the block's
offset), and a children iterator. The children iterator is constructed
depending on the search direction, and with the results of a binary
search over node's children list. It is the only thing that survives a
spilling/push to the stack, everything else is reconstructed. In other
words, each stack spill, will, if the search is still ongoing, cause an
entire re-parsing of the node. Theoretically, this would be a linear
overhead in the number of leaves the search visits. However, one needs
to note:
* the workloads to look for a specific entry are just visiting one leaf,
ever, so this is mostly about workloads that visit larger ranges,
including ones that visit the entire B-tree.
* the requests first hit the page cache, so often the cost is just in
terms of node deserialization
* for nodes that only have leaf nodes as children, no spilling to the
stack-on-heap happens (outside of the initial request where the iterator
is `None`). In other words, for balanced trees, the spilling overhead is
$\Theta\left(\frac{n}{b^2}\right)$, where `b` is the branching factor
and `n` is the number of nodes in the tree. The B-Trees in the current
implementation have a branching factor of roughly `PAGE_SZ/L` where
`PAGE_SZ` is 8192, and `L` is `DELTA_KEY_SIZE = 26` or `KEY_SIZE = 18`
in production code, so this gives us an estimate that we'd be re-loading
an inner node for every 99000 leaves in the B-tree in the worst case.
Due to these points above, I'd say that not fully caching the inner
nodes with inner children is reasonable, especially as we also want to
be fast for the "find one specific entry" workloads, where the stack
content is never accessed: any action to make the spilling
computationally more complex would contribute to wasted cycles here,
even if these workloads "only" spill one node for each depth level of
the b-tree (which is practically always a low single-digit number,
Kleppmann points out on page 81 that for branching factor 500, a four
level B-tree with 4 KB pages can store 250 TB of data).
But disclaimer, this is all stuff I thought about in my head, I have not
confirmed it with any benchmarks or data.
Builds on top of #4863, part of #4743
