Doing to make_multishard_streaming_reader() what the previous commit did
to make_streaming_reader(). In fact, the new compaction_time parameter
is simply forwarded to the make_streaming_reader() on the shard readers.
Call sites are updated, but none opt in just yet.
Opt-in is possible by passing an engaged `compaction_time`
(gc_clock::time_point) to the method. When this new parameter is
disengaged, no compaction happens.
Note that there is a global override, via the
enable_compacting_data_for_streaming_and_repair config item, which can
force-disable this compaction.
Compaction done on the output of the streaming reader does *not*
garbage-collect tombstones!
All call-sites are adjusted (the new parameter is not defaulted), but
none opt in yet. This will be done in separate commit per user.
Compacting can greatly reduce the amount of data to be processed by
streaming and repair, but with certain data shapes, its effectiveness
can be reduced and its CPU overhead might outweight the benefits. This
should very rarely be the case, but leave an off switch in case
this becomes a problem in a deployment.
Not wired yet.
Instead of just a boolean _failed flag, persist the error message of the
exception which caused the repair to fail, and include it in the log
message announcing the failure.
When next_partition() or fast_forward_to() is called. Instead of trying
to simulate a properly closed partition by injecting synthetic mutation
fragments to properly close it.
Currently, the compactor requires a valid stream and thus abandoning a
partition in the middle was not possible. This causes some complications
for the compacting reader, which implements methods such as
`next_partition()` which is possibly called in the middle of a
partition. In this case the compacting reader attempts to close the
partition properly by inserting a synthetic partition-end fragment into
the stream. This is not enough however as it doesn't close any range
tombstone changes that might be active. Instead of piling on more
complexity, add an API to the compactor which allows abandoning the
current partition.
It was found that cached_file dtor can hit the following assert
after OOM
cached_file_test: utils/cached_file.hh:379: cached_file::~cached_file(): Assertion _cache.empty()' failed.`
cached_file's dtor iterates through all entries and evict those
that are linked to LRU, under the assumption that all unused
entries were linked to LRU.
That's partially correct. get_page_ptr() may fetch more than 1
page due to read ahead, but it will only call cached_page::share()
on the first page, the one that will be consumed now.
share() is responsible for automatically placing the page into
LRU once refcount drops to zero.
If the read is aborted midway, before cached_file has a chance
to hit the 2nd page (read ahead) in cache, it will remain there
with refcount 0 and unlinked to LRU, in hope that a subsequent
read will bring it out of that state.
Our main user of cached_file is per-sstable index caching.
If the scenario above happens, and the sstable and its associated
cached_file is destroyed, before the 2nd page is hit, cached_file
will not be able to clear all the cache because some of the
pages are unused and not linked.
A page read ahead will be linked into LRU so it doesn't sit in
memory indefinitely. Also allowing for cached_file dtor to
clear all cache if some of those pages brought in advance
aren't fetched later.
A reproducer was added.
Fixes#14814.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Closes#14818
Fixes https://github.com/scylladb/scylla-docs/issues/4028
The goal of this update is to discourage the use of
the default cassandra superuser in favor of a custom
super user - and explain why it's a good practice.
The scope of this commit:
- Adding a new page on creating a custom superuser.
The page collects and clarifies the information
about the cassandra superuser from other pages.
- Remove the (incomplete) information about
superuser from the Authorization and Authentication
pages, and add the link to the new page instead.
Additionaly, this update will result in better
searchability and ensures language clarity.
Closes#14829
This is to make m.s. initialization more solid and simplify sys.ks.::setup()
Closes#14832
* github.com:scylladb/scylladb:
system_keyspace: Remove unused snitch arg from setup()
messaging_service: Setup preferred IPs from config
Most of the Alternator tests are careful to unconditionally remove the test
tables, even if the test fails. This is important when testing on a shared
database (e.g., DynamoDB) but also useful to make clean shutdown faster
as there should be no user table to flush.
We missed a few such cases in test_gsi.py, and this patch corrects them.
We do this by using the context manager new_test_table() - which
automatically deletes the table when done - instead of the function
create_test_table() which needs an explicit delete at the end.
There are no functional changes in this patch - most of the lines
changed are just reindents.
Signed-off-by: Nadav Har'El <nyh@scylladb.com>
Closes#14835
Make sure _pending_mark_alive_endpoints is unmarked in
any case, including exceptions.
Fixes#14839
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
Closes#14840
Te view updating consumer uses `_buffer_size` to decide when to flush the accumulated mutations, passing them to the actual view building code. This `_buffer_size` is incremented every time a mutation fragment is consumed. This is not exact, as e.g. range tombstones are represented differently in the mutation object, than in the fragment, but it is good enough. There is one flaw however: `_buffer_size` is not incremented when consuming a partition-start fragment. This is when the mutation object is created in the mutation rebuilder. This is not a big problem when partition have many rows, but if the partitions are tiny, the error in accounting quickly becomes significant. If the partitions are empty, `_buffer_size` is not bumped at all for empty partitions, and any number of these can accumulate in the buffer. We have recently seen this causing stalls and OOM as the buffer got to immense size, only containing empty and tiny partitions.
This PR fixes this by accounting the size of the freshly created `mutation` object in `_buffer_size`, after the partition-start fragment is consumed.
Fixes: #14819Closes#14821
* github.com:scylladb/scylladb:
test/boost/view_build_test: add test_view_update_generator_buffering_with_empty_mutations
db/view/view_updating_consumer: account for the size of mutations
mutation/mutation_rebuilder*: return const mutation& from consume_new_partition()
mutation/mutation: add memory_usage()
Population of messageing service preferred IPs cache happens inside
system keyspace setup() call and it needs m.s. per ce and additionally
snitch. Moving preferred ip cache to initial configuration keeps m.s.
start more self-contained and keeps system_keyspace::setup() simpler.
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
Today, test/*/run always kills Scylla at the end of the test with
SIGKILL (kill -9), so the Scylla shutdown code doesn't run. It was
believed that a clean shutdown would take a long time, but in fact,
it turns out that 99% of the shutdown time was a silly sleep in the
gossip code, which this patch disables with the "--shutdown-announce-in-ms"
option.
After enabling this option, clean shutdown takes (in a dev build on
my laptop) just 0.02 seconds. It's worth noting that this shutdown
has no real work to do - no tables to flush, and so on, because the
pytest framework removes all the tables in its own fixture cleanup
phase.
So in this patch, to kill Scylla we use SIGTERM (15) instead of SIGKILL.
We then wait until a timeout of 10 seconds (much much more than 0.02
seconds!) for Scylla to exit. If for some reason it didn't exit (e.g.,
it hung during the shutdown), it is killed again with SIGKILL, which
is guaranteed to succed.
This change gives us two advantages
1. Every test run with test/*/run exercises the shutdown path. It is perhaps
excessive, but since the shutdown is so quick, there is no big downside.
2. In a test-coverage run, a clean shutdown allows flushing the counter
files, which wasn't possible when Scylla was killed with KILL -9.
Fixes#8543
Signed-off-by: Nadav Har'El <nyh@scylladb.com>
Closes#14825
After this series, tablet replication can handle the scenario of bootstrapping new nodes. The ownership is distributed indirectly by the means of a load-balancer which moves tablets around in the background. See docs/dev/topology-over-raft.md for details.
The implementation is by no means meant to be perfect, especially in terms of performance, and will be improved incrementally.
The load balancer will be also kicked by schema changes, so that allocation/deallocation done during table creation/drop will be rebalanced.
Tablet data is streamed using existing `range_streamer`, which is the infrastructure for "the old streaming". This will be later replaced by sstable transfer once integration of tablets with compaction groups is finished. Also, cleanup is not wired yet, also blocked by compaction group integration.
Closes#14601
* github.com:scylladb/scylladb:
tests: test_tablets: Add test for bootstraping a node
storage_service: topology_coordinator: Implement tablet migration state machine
tablets: Introduce tablet_mutation_builder
service: tablet_allocator: Introduce tablet load balancer
tablets: Introduce tablet_map::for_each_tablet()
topology: Introduce get_node()
token_metadata: Add non-const getter of tablet_metadata
storage_service: Notify topology state machine after applying schema change
storage_service: Implement stream_tablet RPC
tablets: Introduce global_tablet_id
stream_transfer_task, multishard_writer: Work with table sharder
tablets: Turn tablet_id into a struct
db: Do not create per-keyspace erm for tablet-based tables
tablets: effective_replication_map: Take transition stage into account when computing replicas
tablets: Store "stage" in transition info
doc: Document tablet migration state machine and load balancer
locator: erm: Make get_endpoints_for_reading() always return read replicas
storage_service: topology_coordinator: Sleep on failure between retries
storage_service: topology_coordinator: Simplify coordinator loop
main: Require experimental raft to enable tablets
A test reproducing #14819, that is, the view update builder not flushing
the buffer when only empty partitions are consumed (with only a
tombstone in them).
All partitions will have a corresponding mutation object in the buffer.
These objects have non-negligible sizes, yet the consumer did not bump
the _buffer_size when a new partition was consumer. This resulted in
empty partitions not moving the _buffer_size at all, and thus they could
accumulate without bounds in the buffer, never triggering a flush just
by themselves. We have recently seen this causing OOM.
This patch fixes that by bumping the _buffer_size with the size of the
freshly created mutation object.
The method is called by db::truncate_table_on_all_shards(), its call-chain, in turn, starts from
- proxy::remote::handle_truncate()
- schema_tables::merge_schema()
- legacy_schema_migrator
- tests
All of the above are easy to get system_keyspace reference from. This, in turn, allows making the method non-static and use query_processor reference from system_keyspace object in stead of global qctx
Closes#14778
* github.com:scylladb/scylladb:
system_keyspace: Make save_truncation_record() non-static
code: Pass sharded<db::system_keyspace>& to database::truncate()
db: Add sharded<system_keyspace>& to legacy_schema_migrator
The function dispatch a background operation that must be
waited on in stop().
Fixes scylladb/scylladb#14791
Signed-off-by: Benny Halevy <bhalevy@scylladb.com>
Closes#14797
Identifies tablet in the scope of the whole cluster. Not to be
confused with tablet replicas, which all share global_tablet_id.
Will be needed by load balancer and tablet migration algorithm to
identify tablets globally.
This erm is not updated when replicating token metadata in
storage_service::replicate_to_all_cores() so will pin token metadata
version and prevent token metadata barrier from finishing.
It is not necessary to have per-keyspace erm for tablet-based tables,
so just don't create it.
It's needed to implement tablet migration. It stores the current step
of tablet migration state machine. The state machine will be advanced
by the topology change coordinator.
See the "Tablet migration" section of topology-over-raft.md
Just a simplification.
Drop the test case from token_metadata which creates pending endpoints
without normal tokens. It fails after this change with exception:
"sorted_tokens is empty in first_token_index!" thrown from
token_metadata::first_token_index(), which is used when calculating
normal endpoints. This test case is not valid, first node inserts
its tokens as normal without going through bootstrap procedure.
When messaging_service shuts down it first sets _shutting_down to true
and proceeds with stopping clients and servers. Stopping clients, in
turn, is calling client.stop() on each.
Setting _shutting_down is used in two places.
First, when a client is stopped it may happen that it's in the middle of
some operation, which may result in call to remove_error_rpc_client()
and not to call .stop() for the second time it just does nothing if the
shutdown flag is set (see 357c91a076).
Second, get_rpc_client() asserts that this flag is not set, so once
shutdown started it can make sure that it will call .stop() on _all_
clients and no new ones would appear in parallel.
However, after shutdown() is complete the _clients vector of maps
remains intact even though all clients from it are stopped. This is not
very debugging-friendly, the clients are better be removed on shutdown.
fixes: #14624
Signed-off-by: Pavel Emelyanov <xemul@scylladb.com>
Closes#14632
It is quite common to stop a tested scylla process with ^C, which will
raise KeyboardInterrupt from subprocess.run(). Catch and swallow this
exception, allowing the post-processing to continue.
The interrupted process has to handle the interrupt correctly too --
flush the coverage data even on premature exit -- but this is for
another patch.
Closes#14815
The new test detected a stack-use-after-return when using table's
as_mutation_source_excluding_staging() for range reads.
This doesn't really affect view updates that generate single
key reads only. So the problem was only stressed in the recently
added test. Otherwise, we'd have seen it when running dtests
(in debug mode) that stress the view update path from staging.
The problem happens because the closure was feeded into
a noncopyable_function that was taken by reference. For range
reads, we defer before subsequent usage of the predicate.
For single key reads, we only defer after finished using
the predicate.
Fix is about using sstable_predicate type, so there won't
be a need to construct a temporary object on stack.
Fixes#14812.
Signed-off-by: Raphael S. Carvalho <raphaelsc@scylladb.com>
Closes#14813
in this series, we use different table names in simple_backlog_controller_test. this test is a test exercising sstables compaction strategies. and it creates and keeps multiple tables in a single test session. but we are going to add metrics on per-table basis, and will use the table's ks and cf as the counter's labels. as the metrics subsystem does not allow multiple counters to share the same label. the test will fail when the metrics are being added.
to address this problem, in this change
1. a new ctor is added for `simple_schema`, so we can create `simple_schema` with different names
2. use the new ctor in simple_backlog_controller_test
Fixes#14767Closes#14783
* github.com:scylladb/scylladb:
test: use different table names in simple_backlog_controller_test
test/lib/simple_schema: add ctor for customizing ks.cf
test/lib/simple_schema: do not hardwire ks.cf
This commit adds the information on how to install ScyllaDB
without root privileges (with "unified installer", but we've
decided to drop that name - see the page title).
The content taken from the website
https://www.scylladb.com/download/?platform=tar&version=scylla-5.2#open-source
is divided into two sections: "Download and Install" and
"Configure and Run ScyllaDB".
In addition, the "Next Steps" section is also copied from
the website, and adjusted to be in sync with other installation
pages in the docs.
Refs https://github.com/scylladb/scylla-docs/issues/4091Closes#14781
In this commit we just pass a fencing_token
through hint_mutation RPC verb.
The hints manager uses either
storage_proxy::send_hint_to_all_replicas or
storage_proxy::send_hint_to_endpoint to send a hint.
Both methods capture the current erm and use the
corresponding fencing token from it in the
mutation or hint_mutation RPC verb. If these
verbs are fenced out, the server stale_topology_exception
is translated to a mutation_write_failure_exception
on the client with an appropriate error message.
The hint manager will attempt to resend the failed
hint from the commitlog segment after a delay.
However, if delivery is unsuccessful, the hint will
be discarded after gc_grace_seconds.
Closes#14580
