for symmetry, let's call it perform_* as it doesn't work like submission
functions which doesn't wait for result, like the one for minor
compaction.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
If memtable flush is segregated into multiple files, partition
estimation becomes innacurate and consequently bloom filters are
bigger than needed, leading to an increase in memory consumption.
To fix this, let's wire adjust_partition_estimate() into the flush
procedure, such that original estimation will be adjusted if
segregation is going to be performed. That's done by feeding
mutation_source_metadata, which will leave original estimation
unchanged if no segregation is needed, but will adjust it
otherwise.
Fixes#9581.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Message-Id: <20211103141600.65806-2-raphaelsc@scylladb.com>
Without tweaking interface, there was no way to adjust estimated
partitions on flush. For example, when segregating a memtable for
TWCS, all produced sstables would have an estimation equal to
the memtable size, even though each only contains a subset of it,
which leads to a significant increase in memory consumption for
bloom filters. Subsequent work will use this interface to perform
the adjustment.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Message-Id: <20211103141600.65806-1-raphaelsc@scylladb.com>
Until now reversed queries were implemented inside
`querier::consume_page` (more precisely, inside the free function
`consume_page` used by `querier::consume_page`) by wrapping the
passed-in reader into `make_reversing_reader` and then consuming
fragments from the resulting reversed reader.
The first couple of commits change that by pushing the reversing down below
the `make_combined_reader` call in `table::query`. This allows
working on improving reversing for memtables independently from
reversing for sstables.
We then extend the `index_reader` with functions that allow
reading the promoted index in reverse.
We introduce `partition_reversing_data_source`, which wraps an sstable data
file and returns data buffers with contents of a single chosen partition
as if the rows were stored in reverse order.
We use the reversing source and the extended index reader in
`mx_sstable_mutation_reader` to implement efficient (at least in theory)
reversed single-partition reads.
The patchset disables cache for reversed reads. Fast-forwarding
is not supported in the mx reader for reversed queries at this point.
Details in commit messages. Read the commits in topological order
for best review experience.
Refs: #9134
(not saying "Fixes" because it's only for single-partition queries
without forwarding)
Closes#9281
* github.com:scylladb/scylla:
table: add option to automatically bypass cache for reversed queries
test: reverse sstable reader with random schema and random mutations
sstables: mx: implement reversed single-partition reads
sstables: mx: introduce partition_reversing_data_source
sstables: index_reader: add support for iterating over clustering ranges in reverse
clustering_key_filter: clustering_key_filter_ranges owning constructor
flat_mutation_reader: mention reversed schema in make_reversing_reader docstring
clustering_key_filter: document clustering_key_filter_ranges::get_ranges
Currently the new reversing sstable algorithms do not support fast
forwarding and the cache does not yet handle reversed results. This
forced us to disable the cache for reversed queries if we want to
guarantee bounded memory. We introduce an option that does this
automatically (without specifying `bypass cache` in the query) and turn
it on by default.
If the user decides that they prefer to keep the cache at the
cost of fetching entire partitions into memory (which may be viable
if their partitions are small) during reversed queries, the option can
be turned off. It is live-updateable.
compaction_info must only contain info data to be exported to the
outside world, whereas compaction_data will contain data for
controlling compaction behavior and stats which change as
compaction progresses.
This separation makes the interface clearer, also allowing for
future improvements like removing direct references to table
in compaction.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
those stats aren't used in compaction stats API and therefore they
can be removed. end_size is added to compaction_result (needed for
updating history) and start_size can be calculated in advance.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Today, compaction is calling compaction manager to register / deregister
the compaction_info created by it.
This is a layer violation because manager sits one layer above
compaction, so manager should be responsible for managing compaction
info.
From now on, compaction_info will be created and managed by
compaction_manager. compaction will only have a reference to info,
which it can use to update the world about compaction progress.
This will allow compaction_manager to be simplified as info can be
coupled with its respective task, allowing duplication to be removed
and layer violation to be fixed.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
There are now 231 translation units that indirectly include commitlog.hh
due to the need to have access to db::commitlog::force_sync.
Move that type to a new file commitlog_types.hh and make it available
without access to the commitlog class.
This reduces the number of translation units that depend on commitlog.hh
to 84, improving compile time.
Currently no mutation-source supports reading in reverse natively but
we are working on changing that, adding native reverse read support to
memtable, cache and sstable readers. To ensure that all mutation
sources work in a correct and uniform manner when reading in reverse,
we add a reverse test to the mutation source test suite. This test
reverses the data that it passes to `populate()`, then reads in
forward order (in reverse compared to the data order). For this we use
the currently established reverse read API: reverse schema (schema
order == query order) and half-reversed (legacy) slice. All mutation
sources are prepared to work with reversed reads, using the
`make_reversing_reader()` adapter. As we progress with our native
reverse support, we will replace these adapters with native reversing
support. As part of this, we push down the reversing reader adapter
currently existing on the `query::consume_page()` level, to the
individual mutation sources.
Closes#9384
* github.com:scylladb/scylla:
test: mutation_reader_test: reversed version of test_clustering_order_merger_sstable_set
querier: consume_page(): remove now unused max_size parameter
test/lib: mutation_source_test: test reading in reverse
test: mutation_reader_test: clustering_combined_reader_mutation_source_test: prepare for reading in reverse
test: flat_mutation_reader_test: test_reverse_reader_is_mutation_source: prepare for reading in reverse
test: mutation_reader_test: test_manual_paused_evictable_reader_is_mutation_source: use query schema instead of table schema
treewide: move reversing to the mutation sources
mutation_query: reconcilable_result_builder: document reverse query preconditions
sstable_set: time_series_sstable_set: reverse mode
mutlishard_mutation_query: set max result size on used permits
db/virtual_table: streaming_virtual_table::as_mutation_source(): use query schema instead of table schema
flat_mutation_reader: make_reversing_reader(): add convenience stored slice
mutation_reader: evictable_reader: add reverse read support
flat_mutation_reader: make_flat_mutation_reader_from_fragments(): add reverse read support
flat_mutation_reader: flat_mutation_reader_from_mutations(): add reverse read support
flat_mutation_reader: flat_mutation_reader_from_mutations(): document preconditions
query-request: introduce `half_reverse_slice`
flat_mutation_reader_assertions: log what's expected
Currently the following can happen:
1) there's ongoing compaction with input sstable A, so sstable set
and backlog tracker both contains A.
2) ongoing compaction replaces input sstable A by B, so sstable set
contains only B now.
3) schema is updated, so a new backlog tracker is built without A
because sstable set now contains only B.
4) ongoing compaction tries to remove A from tracker, but it was
excluded in step 3.
5) tracker can now have a negative value if table is decreasing in
size, which leads to log(<negative number>) == -NaN
This problem happens because backlog tracker updates are decoupled
from sstable set updates. Given that the essential content of
backlog tracker should be the same as one of sstable set, let's move
tracker management to table.
Whenever sstable set is updated, backlog tracker will be updated with
the same changes, making their management less error prone.
Fixes#9157
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
The generic backlog formula is: ALL + PARTIAL - COMPACTING
With transfer_ongoing_charges() we already ignore the effect of
ongoing compactions on COMPACTING as we judge them to be pointless.
But ongoing compactions will run to completion, meaning that output
sstables will be added to ALL anyway, in the formula above.
With stop_tracking_ongoing_compactions(), input sstables are never
removed from the tracker, but output sstables are added, which means
we end up with duplicate backlog in the tracker.
By removing this tracking mechanism, pointless ongoing compaction
will be ignored as expected and the leaks will be fixed.
Later, the intention is to force a stop on ongoing compactions if
strategy has changed as they're pointless anyway.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
"This series removes layer violation in compaction, and also
simplifies compaction manager and how it interacts with compaction
procedure."
* 'compaction_manager_layer_violation_fix/v3' of github.com:raphaelsc/scylla:
compaction: split compaction info and data for control
compaction_manager: use task when stopping a given compaction type
compaction: remove start_size and end_size from compaction_info
compaction_manager: introduce helpers for task
compaction_manager: introduce explicit ctor for task
compaction: kill sstables field in compaction_info
compaction: kill table pointer in compaction_info
compaction: simplify procedure to stop ongoing compactions
compaction: move management of compaction_info to compaction_manager
compaction: move output run id from compaction_info into task
compaction_info must only contain info data to be exported to the
outside world, whereas compaction_data will contain data for
controlling compaction behavior and stats which change as
compaction progresses.
This separation makes the interface clearer, also allowing for
future improvements like removing direct references to table
in compaction.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
those stats aren't used in compaction stats API and therefore they
can be removed. end_size is added to compaction_result (needed for
updating history) and start_size can be calculated in advance.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Today, compaction is calling compaction manager to register / deregister
the compaction_info created by it.
This is a layer violation because manager sits one layer above
compaction, so manager should be responsible for managing compaction
info.
From now on, compaction_info will be created and managed by
compaction_manager. compaction will only have a reference to info,
which it can use to update the world about compaction progress.
This will allow compaction_manager to be simplified as info can be
coupled with its respective task, allowing duplication to be removed
and layer violation to be fixed.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Add the content of the compaction stats introduced in the previous patch
to the tracing data. This will help diagnose query performance related
problems caused by tombstones.
Currently the following can happen:
1) there's ongoing compaction with input sstable A, so sstable set
and backlog tracker both contains A.
2) ongoing compaction replaces input sstable A by B, so sstable set
contains only B now.
3) schema is updated, so a new backlog tracker is built without A
because sstable set now contains only B.
4) ongoing compaction tries to remove A from tracker, but it was
excluded in step 3.
5) tracker can now have a negative value if table is decreasing in
size, which leads to log(<negative number>) == -NaN
This problem happens because backlog tracker updates are decoupled
from sstable set updates. Given that the essential content of
backlog tracker should be the same as one of sstable set, let's move
tracker management to table.
Whenever sstable set is updated, backlog tracker will be updated with
the same changes, making their management less error prone.
Fixes#9157
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
The generic back formula is: ALL + PARTIAL - COMPACTING
With transfer_ongoing_charges() we already ignore the effect of
ongoing compactions on COMPACTING as we judge them to be pointless.
But ongoing compactions will run to completion, meaning that output
sstables will be added to ALL anyway, in the formula above.
With stop_tracking_ongoing_compactions(), input sstables are never
removed from the tracker, but output sstables are added, which means
we end up with duplicate backlog in the tracker.
By removing this tracking mechanism, pointless ongoing compaction
will be ignored as expected and the leaks will be fixed.
Later, the intention is to force a stop on ongoing compactions if
strategy has changed as they're pointless anyway.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
The semaphore inside was never accessed and `max_memory_for_unlimited_query`
was always equal to `*cmd.max_result_size` so the parameter was completely
redundant.
`cmd.max_result_size` is supposed to be always set in the affected
functions - which are executed on the replica side - as soon as the
replica receives the `read_command` object, in case the parameter was
not set by the coordinator. However, we don't have a guarantee at the
type level (it's still an `optional`). Many places used
`*cmd.max_result_size` without even an assertion.
We make the code a bit safer, we check for `cmd.max_result_size` and if
it's indeed engaged, store it in `reader_permit`. We then access it from
`reader_permit` where necessary. If `cmd.max_result_size` is not set, we
assume this is an unlimited query and obtain the limit from
`get_unlimited_query_max_result_size`.
We define the native reverse format as a reversed mutation fragment
stream that is identical to one that would be emitted by a table with
the same schema but with reversed clustering order. The main difference
to the current format is how range tombstones are handled: instead of
looking at their start or end bound depending on the order, we always
use them as-usual and the reversing reader swaps their bounds to
facilitate this. This allows us to treat reversed streams completely
transparently: just pass along them a reversed schema and all the
reader, compacting and result building code is happily ignorant about
the fact that it is a reversed stream.
The reader is used by load and stream to read sstables from the upload
directory which are not guaranteed to belong to the local shard.
Using the make_range_sstable_reader instead of
make_local_shard_sstable_reader.
Tests:
backup_restore_tests.py:TestBackupRestore.load_and_stream_using_snapshot_test
backup_restore_tests.py:TestBackupRestore.load_and_stream_to_new_cluster_2_test
backup_restore_tests.py:TestBackupRestore.load_and_stream_to_new_cluster_1_test
migration_test.py:TestLoadAndStream.load_and_stream_asymmetric_cluster_test
migration_test.py:TestLoadAndStream.load_and_stream_decrease_cluster_test
migration_test.py:TestLoadAndStream.load_and_stream_frozen_pk_test
migration_test.py:TestLoadAndStream.load_and_stream_increase_cluster_test
migration_test.py:TestLoadAndStream.load_and_stream_primary_replica_only_test
Fixes#9173Closes#9185
The table::run_compaction is a trivial wrapper for
table::compact_sstables.
We have lots of similar {start, trigger, run}_compaction functions.
Dropping the run_compaction wrapper to reduce confusion.
Closes#9161
NOTE: this series depends on a Seastar submodule update, currently queued in next: 0ed35c6af052ab291a69af98b5c13e023470cba3
In order to avoid needless throwing, exceptions are passed
directly wherever possible. Two mechanisms which help with that are:
1. `make_exception_future<>` for futures
2. `co_return coroutine::exception(...)` for coroutines
which return `future<T>` (the mechanism does not work for `future<>`
without parameters, unfortunately)
Tests: unit(release)
Closes#9079
* github.com:scylladb/scylla:
system_keyspace: pass exceptions without throwing
sstables: pass exceptions without throwing
storage_proxy: pass exceptions without throwing
multishard_mutation_query: pass exceptions without throwing
client_state: pass exceptions without throwing
flat_mutation_reader: pass exceptions without throwing
table: pass exceptions without throwing
commitlog: pass exceptions without throwing
compaction: pass exceptions without throwing
database: pass exceptions without throwing
Continuing the work from e4eb7df1a1, let's guarantee
serialization of sstable set updates by making all sites acquire
the mutation permit. Then table no longer rely on serialization
mechanism of row cache's update functions.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Message-Id: <20210728174740.78826-1-raphaelsc@scylladb.com>
In order to avoid needless throwing, exceptions are passed
directly wherever possible. Two mechanisms which help with that are:
1. make_exception_future<> for futures
2. co_return coroutine::exception(...) for coroutines
which return future<T> (the mechanism does not work for future<>
without parameters, unfortunately)
Today, table relies on row_cache::invalidate() serialization for
concurrent sstable list updates to produce correct results.
That's very error prone because table is relying on an implementation
detail of invalidate() to get things right.
Instead, let's make table itself take care of serialization on
concurrent updates.
To achieve that, sstable_list_builder is introduced. Only one
builder can be alive for a given table, so serialization is guaranteed
as long as the builder is kept alive throughout the update procedure.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Message-Id: <20210721001716.210281-1-raphaelsc@scylladb.com>
That function is dangerously used by distributed loader, as the latter
was responsible for invalidating cache for new sstable.
load_sstable() is an unsafe alternative to
add_sstable_and_update_cache() that should never have been used by
the outside world. Instead, let's kill it and make loader use
the safe alternative instead.
This will also make it easier to make sure that all concurrent updates
to sstable set are properly serialized.
Additionally, this may potentially reduce the amount of data evicted
from the cache, when the sstables being imported have a narrow range,
like high level sstables imported from a LCS table. Unlikely but
possible.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Message-Id: <20210721131949.26899-1-raphaelsc@scylladb.com>
In order to avoid data loss bugs, that could come due to lack of
serialization when using the preemptable build_new_sstable_list(),
let's document the serialization requirement.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Message-Id: <20210714201301.188622-1-raphaelsc@scylladb.com>
This patchset combines two important changes to the way reader permits
are created and admitted:
1) It switches admission to be up-front.
2) It changes the admission algorithm.
(1) Currently permits are created before the read is started, but they
only wait for admission when going to the disk. This leaves the
resources consumption of cache and memtables reads unbounded, possibly
leading to OOM (rare but happens). This series changes this that permits
are admitted at the moment they are creating making admission up-front
-- at least those reads that pass admission at all (some don't).
(2) Admission currently is based on availability of resources. We have a
certain amount of memory available, which derived from the memory
available to the shard, as well a hardcoded count resource. Reads are
admitted when a count and a certain amount (base cost) of memory is
available. This patchset adds a new aspect to this admission process
beyond the existing resource availability: the number of used/blocked
reads. Namely it only admits new reads if in addition to the necessary
amount of resources being available, all currently used readers are
blocked. In other words we only admit new reads if all currently
admitted reads requires something other than CPU to progress. They are
either waiting on I/O, a remote shard, or attention from their consumers
(not used currently).
The reason for making these two changes at the same time is that
up-front admission means cache reads now need to obtain a permit too.
For cache reads the optimal concurrency is 1. Anything above that just
increases latency (without increasing throughput). So we want to make sure
that if a cache reader hits it doesn't get any competition for CPU and
it can run to completion. We admit new reads only if the read misses and
has to go to disk.
A side effect of these changes is that the execution stages from the
replica-side read path are replaced with the reader concurrency
semaphore as an execution stage. This is necessary due to bad
interaction between said execution stages and up-front admission. This
has an important consequence: read timeouts are more strictly enforced
because the execution stage doesn't have a timeout so it can execute
already timed-out reads too. This is not the case with the semaphore's
queue which will drop timed-out reads. Another consequence is that, now
data and mutation reads share the same execution stage, which increases
its effectiveness, on the other hand system and user reads don't
anymore.
Fixes: #4758Fixes: #5718
Tests: unit(dev, release, debug)
* 'reader-concurrency-semaphore-in-front-of-the-cache/v5.3' of https://github.com/denesb/scylla: (54 commits)
test/boost/reader_concurrency_semaphore_test: add used/blocked test
test/boost/reader_concurrency_semaphore_test: add admission test
reader_permit: add operator<< for reader_resources
reader_concurrency_semaphore: add reads_{admitted,enqueued} stats
table: make_sstable_reader(): fix indentation
table: clean up make_sstable_reader()
database: remove now unused query execution stages
mutation_reader: remove now unused restricting_reader
sstables: sstable_set: remove now unused make_restricted_range_sstable_reader()
reader_permit: remove now unused wait_admission()
reader_concurrency_semaphore: remove now unused obtain_permit_nowait()
reader_concurrency_semaphore: admission: flip the switch
database: increase semaphore max queue size
test: index_with_paging_test: increase semaphore's queue size
reader_concurrency_semaphore: add set_max_queue_size()
test: mutation_reader_test: remove restricted reader tests
reader_concurrency_semaphore: remove now unused make_permit()
test: reader_concurrency_semaphore_test: move away from make_permit()
test: move away from make_permit()
treewide: use make_tracking_only_permit()
...
This patch flips two "switches":
1) It switches admission to be up-front.
2) It changes the admission algorithm.
(1) by now all permits are obtained up-front, so this patch just yanks
out the restricted reader from all reader stacks and simultaneously
switches all `obtain_permit_nowait()` calls to `obtain_permit()`. By
doing this admission is now waited on when creating the permit.
(2) we switch to an admission algorithm that adds a new aspect to the
existing resource availability: the number of used/blocked reads. Namely
it only admits new reads if in addition to the necessary amount of
resources being available, all currently used readers are blocked. In
other words we only admit new reads if all currently admitted reads
requires something other than CPU to progress. They are either waiting
on I/O, a remote shard, or attention from their consumers (not used
currently).
We flip these two switches at the same time because up-front admission
means cache reads now need to obtain a permit too. For cache reads the
optimal concurrency is 1. Anything above that just increases latency
(without increasing throughput). So we want to make sure that if a cache
reader hits it doesn't get any competition for CPU and it can run to
completion. We admit new reads only if the read misses and has to go to
disk.
Another change made to accommodate this switch is the replacement of the
replica side read execution stages which the reader concurrency
semaphore as an execution stage. This replacement is needed because with
the introduction of up-front admission, reads are not independent of
each other any-more. One read executed can influence whether later reads
executed will be admitted or not, and execution stages require
independent operations to work well. By moving the execution stage into
the semaphore, we have an execution stage which is in control of both
admission and running the operations in batches, avoiding the bad
interaction between the two.
