It was observed that some use cases might append old data constantly to
memtable, blocking GC of expired tombstones.
That's because timestamp of memtable is unconditionally used for
calculating max purgeable, even when the memtable doesn't contain the
key of the tombstone we're trying to GC.
The idea is to treat memtable as we treat L0 sstables, i.e. it will
only prevent GC if it contains data that is possibly shadowed by the
expired tombstone (after checking for key presence and timestamp).
Memtable will usually have a small subset of keys in largest tier,
so after this change, a large fraction of keys containing expired
tombstones can be GCed when memtable contains old data.
Fixes#17599.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
(cherry picked from commit 38699f6c3d)
Closesscylladb/scylladb#19551
TWCS off-strategy suffers with 100% space overhead, so a big TWCS table
can cause scylla to run out of disk space during node ops.
To not penalize TWCS tables, that take a small percentage of disk,
with increased write ampl, TWCS off-strategy will be restricted to
10% of free disk space. Then small tables can still compact all
disjoint sstables in a single round.
Fixes#16514.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
(cherry picked from commit ace4e5111e)
After this, TWCS reshape procedure can be changed to limit job
to 10% of available space.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
(cherry picked from commit 0ce8ee03f1)
When a compaction strategy uses garbage collected sstables to track
expired tombstones, do not use complete partition estimates for them,
instead, use a fraction of it based on the droppable tombstone ratio
estimate.
Fixes#18283
Signed-off-by: Lakshmi Narayanan Sreethar <lakshmi.sreethar@scylladb.com>
Closesscylladb/scylladb#18465
(cherry picked from commit d39adf6438)
Closesscylladb/scylladb#18656
Since commit f1bbf70, many compaction types can do cleanup work, but turns out
we forgot to invalidate cache on their completion.
So if a node regains ownership of token that had partition deleted in its previous
owner (and tombstone is already gone), data can be resurrected.
Tablet is not affected, as it explicitly invalidates cache during migration
cleanup stage.
Scylla 5.4 is affected.
Fixes#17501.
Fixes#17452.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Closesscylladb/scylladb#17502
(cherry picked from commit f07c233ad5)
Regular compaction tasks are internal.
Adjust test_compaction_task accordingly: modify test_regular_compaction_task,
delete test_running_compaction_task_abort (relying on regular compaction)
which checks are already achived by test_not_created_compaction_task_abort.
Rename the latter.
(cherry picked from commit 6b87778ef2)
For major compacting all tables in the database.
The advantage of this api is that `commitlog->force_new_active_segment`
happens only once in `database::flush_all_tables` rather than
once per keyspace (when `nodetool compact` translates to
a sequence of `/storage_service/keyspace_compaction` calls).
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
(cherry picked from commit b12b142232)
Major compaction already flushes each table to make
sure it considers any mutations that are present in the
memtable for the purpose of tombstone purging.
See 64ec1c6ec6
However, tombstone purging may be inhibited by data
in commitlog segments based on `gc_time_min` in the
`tombstone_gc_state` (See f42eb4d1ce).
Flushing all sstables in the database release
all references to commitlog segments and there
it maximizes the potential for tombstone purging,
which is typically the reason for running major compaction.
However, flushing all tables too frequently might
result in tiny sstables. Since when flushing all
keyspaces using `nodetool flush` the `force_keyspace_compaction`
api is invoked for keyspace successively, we need a mechanism
to prevent too frequent flushes by major compaction.
Hence a `compaction_flush_all_tables_before_major_seconds` interval
configuration option is added (defaults to 24 hours).
In the case that not all tables are flushed prior
to major compaction, we revert to the old behavior of
flushing each table in the keyspace before major-compacting it.
Fixesscylladb/scylladb#15777
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
(cherry picked from commit 66ba983fe0)
When flushing is done externally, e.g. by running
`nodetool flush` prior to `nodetool compact`,
flush_memtables=false can be passed to skip flushing
of tables right before they are major-compacted.
This is useful to prevent creation of small sstables
due to excessive memtable flushing.
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
(cherry picked from commit 1fd85bd37b)
TWCS tables require partition estimation adjustment as incoming streaming data can be segregated into the time windows.
Turns out we had two problems in this area that leads to suboptimal bloom filters.
1) With off-strategy enabled, data segregation is postponed, but partition estimation was adjusted as if segregation wasn't postponed. Solved by not adjusting estimation if segregation is postponed.
2) With off-strategy disabled, data segregation is not postponed, but streaming didn't feed any metadata into partition estimation procedure, meaning it had to assume the max windows input data can be segregated into (100). Solved by using schema's default TTL for a precise estimation of window count.
For the future, we want to dynamically size filters (see https://github.com/scylladb/scylladb/issues/2024), especially for TWCS that might have SSTables that are left uncompacted until they're fully expired, meaning that the system won't heal itself in a timely manner through compaction on a SSTable that had partition estimation really wrong.
Fixes https://github.com/scylladb/scylladb/issues/15704.
Closesscylladb/scylladb#15938
* github.com:scylladb/scylladb:
streaming: Improve partition estimation with TWCS
streaming: Don't adjust partition estimate if segregation is postponed
(cherry picked from commit 64d1d5cf62)
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Closesscylladb/scylladb#16671
The polling loop was intended to ignore
`condition_variable_timed_out` and check for progress
using a longer `max_idle_duration` timeout in the loop.
Fixes#15669
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
Closesscylladb/scylladb#15671
(cherry picked from commit 68a7bbe582)
After "repair: Get rid of the gc_grace_seconds", the sstable's schema (mode,
gc period if applicable, etc) is used to estimate the amount of droppable
data (or determine full expiration = max_deletion_time < gc_before).
It could happen that the user switched from timeout to repair mode, but
sstables will still use the old mode, despite the user asked for a new one.
Another example is when you play with value of grace period, to prevent
data resurrection if repair won't be able to run in a timely manner.
The problem persists until all sstables using old GC settings are recompacted
or node is restarted.
To fix this, we have to feed latest schema into sstable procedures used
for expiration purposes.
Fixes#15643.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Closesscylladb/scylladb#15746
(cherry picked from commit fded314e46)
The estimation assumes that size of other components are irrelevant,
when estimating the number of partitions for each output sstable.
The sstables are split according to the data file size, therefore
size of other files are irrelevant for the estimation.
With certain data models, like single-row partitions containing small
values, the index could be even larger than data.
For example, assume index is as large as data, then the estimation
would say that 2x more sstables will be generated, and as a result,
each sstable are underestimated to have 2x less keys.
Fix it by only accounting size of data file.
Fixes#15726.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Closesscylladb/scylladb#15727
Before integration with task manager the state of one shard repair
was kept in repair_info. repair_info object was destroyed immediately
after shard repair was finished.
In an integration process repair_info's fields were moved to
shard_repair_task_impl as the two served the similar purposes.
Though, shard_repair_task_impl isn't immediately destoyed, but is
kept in task manager for task_ttl seconds after it's complete.
Thus, some of repair_info's fields have their lifetime prolonged,
which makes the repair state change delayed.
Release shard_repair_task_impl resources immediately after shard
repair is finished.
Fixes: #15505.
Closesscylladb/scylladb#15506
SSTable runs work hard to keep the disjointness invariant, therefore they're
expensive to build from scratch.
For every insertion, it keeps the elements sorted by their first key in
order to reject insertion of element that would introduce overlapping.
Additionally, a sstable run can grow to dozens of elements (or hundreds)
therefore, we can also make interaction with compaction strategies more
efficient by not copying them when building a list of candidates in compaction
manager. And less fragile by filtering out any sstable runs that are not
completely eligible for compaction.
Previously, ICS had to give up on using runs managed by sstable set due to
fragility of the interface (meaning runs are being built from scratch
on every call to the strategy, which is very inefficient, but that had to
be done for correctness), but now we can restore that.
Closesscylladb/scylladb#15440
* github.com:scylladb/scylladb:
compaction: Switch to strategy_control::candidates() for regular compaction
tests: Prepare sstable_compaction_test for change in compaction_strategy interface
compaction: Allow strategy to retrieve candidates either as sstables or runs
compaction: Make get_candidates() work with frozen_sstable_run too
sstables: add sstable_run::run_identifier()
sstables: tag sstable_run::insert() with nodiscard
sstables: Make all_sstable_runs() more efficient by exposing frozen shared runs
sstables: Simplify sstable_set interface to retrieve runs
Most of the time only the roots of tasks tree should be non internal.
Change default implementation of is_internal and delete overrides
consistent with it.
Closesscylladb/scylladb#15353
Now everything is prepared for the switch, let's do it.
Now let's wait for ICS to enjoy the set of changes.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
That's needed for upcoming changes that will allow ICS to efficiently
retrieve sstable runs.
Next patch will remove candidates from compaction_strategy's interface
to retrieve candidates using this one instead.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Off-strategy suffers with a 100% space overhead, as it adopted
a sort of all or nothing approach. Meaning all input sstables,
living in maintenance set, are kept alive until they're all
reshaped according to the strategy criteria.
Input sstables in off-strategy are very likely to be mostly disjoint,
so it can greatly benefit from incremental compaction.
The incremental compaction approach is not only good for
decreasing disk usage, but also memory usage (as metadata of
input and output live in memory), and file desc count, which
takes memory away from OS.
Turns out that this approach also greatly simplifies the
off-strategy impl in compaction manager, as it no longer have
to maintain new unused sstables and mark them for
deletion on failure, and also unlink intermediary sstables
used between reshape rounds.
Fixes https://github.com/scylladb/scylladb/issues/14992.
Closesscylladb/scylladb#15400
* github.com:scylladb/scylladb:
test: Verify that off-strategy can do incremental compaction
compaction: Clear pending_replacement list when tombstone GC is disabled
compaction: Enable incremental compaction on off-strategy
compaction: Extend reshape type to allow for incremental compaction
compaction: Move reshape_compaction in the source
compaction: Enable incremental compaction only if replacer callback is engaged
pending_replacement list is used by incremental compaction to
communicate to other ongoing compactions about exhausted sstables
that must be replaced in the sstable set they keep for tombstone
GC purposes.
Reshape doesn't enable tombstone GC, so that list will not
be cleared, which prevents incremental compaction from releasing
sstables referenced by that list. It's not a problem until now
where we want reshape to do incremental compaction.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Off-strategy suffers with a 100% space overhead, as it adopted
a sort of all or nothing approach. Meaning all input sstables,
living in maintenance set, are kept alive until they're all
reshaped according to the strategy criteria.
Input sstables in off-strategy are very likely to mostly disjoint,
so it can greatly benefit from incremental compaction.
The incremental compaction approach is not only good for
decreasing disk usage, but also memory usage (as metadata of
input and output live in memory), and file desc count, which
takes memory away from OS.
Turns out that this approach also greatly simplifies the
off-strategy impl in compaction manager, as it no longer have
to maintain new unused sstables and mark them for
deletion on failure, and also unlink intermediary sstables
used between reshape rounds.
Fixes#14992.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
That's done by inheriting regular_compaction, which implement
incremental compaction. But reshape still implements its own
methods for creating writer and reader. One reason is that
reshape is not driven by controller, as input sstables to it
live in maintenance set. Another reason is customization
of things like sstable origin, etc.
stop_sstable_writer() is extended because that's used by
regular_compaction to check for possibility of removing
exhausted sstables earlier whenever an output sstable
is sealed.
Also, incremental compaction will be unconditionally
enabled for ICS/LCS during off-strategy.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
That's in preparation to next change that will make reshape
inherit from regular compaction.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Currently, the moved-object's manager pointer is moved into the
constructed object, but without fixing the registration to
point to the moved-to object, causing #15248.
Although we could properly move the registration from
the moved-from object to the moved-to one, it is simpler
to just disallow moving a registered tracker, since it's
not needed anywhere. This way we just don't need to mess
with the trackers' registration.
The move-assignment operator has a similar problem,
therefore it is deleted in this series, and the function is
renamed to `transfer_backlog` that just doesn't deal with the
moved-from registration. This is safe since it's only used internally
by the compaction manager.
Fixes#15248Closesscylladb/scylladb#15445
* github.com:scylladb/scylladb:
compaction_state: store backlog_track in std::optional
compaction_backlog_tracker: do not allow moving registered trackers
Compaction tasks executors serve two different purposes - as compaction
manager related entity they execute compaction operation and as task
manager related entity they track compaction status.
When one role depends on the other, as it currently is for
compaction_task_impl::done() and compaction_task_executor::compaction_done(),
requirements of both roles need to be satisfied at the same time in each
corner case. Such complexity leads to bugs.
To prevent it, compaction_task_impl::done() of executors no longer depends
on compaction_task_executor::compaction_done().
Fixes: #14912.
Closesscylladb/scylladb#15140
* github.com:scylladb/scylladb:
compaction: warn about compaction_done()
compaction: do not run stopped compaction
compaction: modify lowest compaction tasks' run method
compaction: pass do_throw_if_stopping to compaction_task_executor
So that replacing it will destroy the previous tracker
and unregister it before assigning the new one and
then registering it.
This is safer than assiging it in place.
With that, the move assignment operator is not longer
used and can be deleted.
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
Currently, the moved-object's manager pointer is moved into the
constructed object, but without fixing the registration to
point to the moved-to object, causing #15248.
Although we could properly move the registration from
the moved-from object to the moved-to one, it is simpler
to just disallow moving a registered tracker, since it's
not needed anywhere. This way we just don't need to mess
with the trackers' registration.
With that in mind, when move-assigning a compaction_backlog_tracker
the existing tracker can remain registered.
Fixesscylladb/scylladb#15248
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
That's needed for enabling incremental compaction to operate, and
needed for subsequent work that enables incremental compaction
for off-strategy, which in turn uses reshape compaction type.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Split compaction_strategy_impl constructor into methods that will
be reused for validation.
Add additional checks providing that options' values are legal.
Add compaction_strategy_impl::validate_min_max_threshold method
that will be used to validate min and max threshold values
for different compaction methods.
Split size_tiered_compaction_strategy_options constructor into
methods that will be reused for validation.
Add additional checks providing that options' values are legal.
To be consistent with other compaction_strategy_options,
time_window_compaction_strategy_options uses compaction_strategy_impl::get_value
and cql3::statements::property_definitions::to_long helpers for
parsing.
Add temporarily empty validate method to compaction_strategy_options.
The method will validate the options and help determining whether
only the allowed options were set.
compaction_done() returns ready future before compaction_task_executor::run_compaction()
even though the compaction did not start.
Make compaction_done() private and add a comment to warn against
incorrect usage.
Before compaction_task_executor::do_run is called, the executor can
be already aborted. Check if compaction was stopped and set
_compaction_done to exceptional future.
For compaction_task_executors, unlike for all other task manager
tasks, run method does not embrace operations performed in a scope
of a task, but only waits until shared_future connected with
the operations is resolved.
Apart from breaking task manager task conventions, such a run method
must consider all corner cases, not to break task manager or
compaction manager functionality.
To fix existing and prevent further bugs related to task manager
and compaction manager coexistence, call perform_task inside
run method and wait for it in a standard way.
Executors that are not going to be reflected in task manager run call
perform_task the old way.
Make sure the compaction_state:s are idle before
they are destroyed. Although all tasks are stopped
in stop_ongoing_compactions, make sure there is
fiber holding the compaction_state gate.
compaction_manager::remove now needs to close the
compaction_state gate and to stop_ongoing_compactions
only if the gate is not closed yet.
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
