7e2875d648
The goal of merge is to reduce the tablet count for a shrinking table. Similar to how split increases the count while the table is growing. The load balancer decision to merge is implemented today (came with infrastructure introduced for split), but it wasn't handled until now. Initial tablet count is respected while the table is in "growing mode". For example, the table leaves it if there was a need to split above the initial tablet count. After the table leaves the mode, the average size can be trusted to determine that the table is shrinking. Merge decision is emitted if the average tablet size is 50% of the target. Hysteresis is applied to avoid oscillations between split and merges. Similar to split, the decision to merge is recorded in tablet map's resize_type field with the string "merge". This is important in case of coordinator failover, so new coordinator continues from where the old left off. Unlike split, the preparation phase during merge is not done by the replica (with split compactions), but rather by the coordinator by co-locating sibling tablets in the same node's shard. We can define sibling tablets as tablets that have contiguous range and will become one after merge. The concept is based on the power-of-two constraint and token contiguity. For example, in a table with 4 tablets, tablets of ids 0 and 1 are siblings, 2 and 3 are also siblings. The algorithm for co-locating sibling tablets is very simple. The balancer is responsible for it, and it will emit migrations so that "odd" tablet will follow the "even" one. For example, tablet 1 will be migrated to where tablet 0 lives. Co-location is low in priority, it's not the end of the world to delay merge, but it's not ideal to delay e.g. decommission or even regular load balancing as that can translate into temporary unbalancing, impacting the user activities. So co-location migrations will happen when there is no more important work to do. While regular balancing is higher in priority, it will not undo the co-location work done so far. It does that by treating co-located tablets as if they were already merged. The load inversion convergence check was adjusted so balancer understand when two tablets are being migrated instead of one, to avoid oscillations. When balancer completes co-location work for a table undergoing merge, it will put the id of the table into the resize_plan, which is about communicating with the topology coordinator that a table is ready for it. With all sibling tablets co-located, the coordinator can resize the tablet map (reduce it by a factor of 2) and record the new map into group0. All the replicas will react to it (on token metadata update) by merging the storage (memtable(s) + sstables) of sibling tablets into one. Fixes #18181. system test details: test: https://github.com/pehala/scylla-cluster-tests/blob/tablets_split_merge/tablets_split_merge_test.py yaml file: https://github.com/pehala/scylla-cluster-tests/blob/tablets_split_merge/test-cases/features/tablets/tablets-split-merge-test.yaml instance type: i3.8xlarge nodes: 3 target tablet size: 0.5G (scaled down by 10, to make it easier to trigger splits and merges) description: multiple cycles of growing and shrinking the data set in order to trigger splits and merges. data_set_size: ~100G initial_tablets: 64, so it grew to 128 tablets on split, and back to 64 on merge. latency of reads and writes that happened in parallel to split and merge: ``` $ for i in scylla-bench*; do cat $i | grep "Mode\|99th:\|99\.9th:"; done Mode: write 99.9th: 3.145727ms 99th: 1.998847ms 99.9th: 3.145727ms 99th: 2.031615ms Mode: read 99.9th: 3.145727ms 99th: 2.031615ms 99.9th: 3.145727ms 99th: 2.031615ms Mode: write 99.9th: 3.047423ms 99th: 1.933311ms 99.9th: 3.047423ms 99th: 1.933311ms Mode: read 99.9th: 3.145727ms 99th: 1.900543ms 99.9th: 3.145727ms 99th: 1.900543ms Mode: write 99.9th: 5.079039ms 99th: 3.604479ms 99.9th: 35.389439ms 99th: 25.624575ms Mode: write 99.9th: 3.047423ms 99th: 1.998847ms 99.9th: 3.047423ms 99th: 1.998847ms Mode: read 99.9th: 3.080191ms 99th: 2.031615ms 99.9th: 3.112959ms 99th: 2.031615ms ``` Closes scylladb/scylladb#20572 * github.com:scylladb/scylladb: docs: Document tablet merging tests/boost: Add test to verify correctness of balancer decisions during merge tests/topology_experimental_raft: Add tablet merge test service: Handle exception when retrying split service: Co-locate sibling tablets for a table undergoing merge gms: Add cluster feature for tablet merge service: Make merge of resize plan commutative replica: Implement merging of compaction groups on merge completion replica: Handle tablet merge completion service: Implement tablet map resize for merge locator: Introduce merge_tablet_info() service: Rename topology::transition_state::tablet_split_finalization service: Respect initial_tablet_count if table is in growing mode service: Wire migration_tablet_set into the load balancer locator: Add tablet_map::sibling_tablets() service: Introduce sorted_replicas_for_tablet_load() locator/tablets: Extend tablet_replica equality comparator to three-way service: Introduce alias to per-table candidate map type service: Add replication constraint check variant for migration_tablet_set service: Add convergence check variant for migration_tablet_set service: Add migration helpers for migration_tablet_set service/tablet_allocator: Introduce migration_tablet_set service: Introduce migration_plan::add(migrations_vector) locator/tablets: Introduce tablet_map::for_each_sibling_tablets() locator/tablets: Introduce tablet_map::needs_merge() locator/tablets: Introduce resize_decision::initial_decision() locator/tablets: Fix return type of three-way comparison operators service: Extract update of node load on migrations service: Extract converge check for intra-node migration service: Extract erase of tablet replicas from candidate list scripts/tablet-mon: Allow visualization of tablet id
Scylla unit tests using C++ and the Boost test framework
The source files in this directory are Scylla unit tests written in C++ using the Boost.Test framework. These unit tests come in three flavors:
-
Some simple tests that check stand-alone C++ functions or classes use Boost's
BOOST_AUTO_TEST_CASE. -
Some tests require Seastar features, and need to be declared with Seastar's extensions to Boost.Test, namely
SEASTAR_TEST_CASE. -
Even more elaborate tests require not just a functioning Seastar environment but also a complete (or partial) Scylla environment. Those tests use the
do_with_cql_env()ordo_with_cql_env_thread()function to set up a mostly-functioning environment behaving like a single-node Scylla, in which the test can run.
While we have many tests of the third flavor, writing new tests of this type should be reserved to white box tests - tests where it is necessary to inspect or control Scylla internals that do not have user-facing APIs such as CQL. In contrast, black-box tests - tests that can be written only using user-facing APIs, should be written in one of newer test frameworks that we offer - such as test/cqlpy or test/alternator (in Python, using the CQL or DynamoDB APIs respectively) or test/cql (using textual CQL commands), or - if more than one Scylla node is needed for a test - using the test/topology* framework.
Running tests
Because these are C++ tests, they need to be compiled before running.
To compile a single test executable row_cache_test, use a command like
ninja build/dev/test/boost/row_cache_test
You can also use ninja dev-test to build all C++ tests, or use
ninja deb-build to build the C++ tests and also the full Scylla executable
(however, note that full Scylla executable isn't needed to run Boost tests).
Replace "dev" by "debug" or "release" in the examples above and below to use the "debug" build mode (which, importantly, compiles the test with ASAN and UBSAN enabling on and helps catch difficult-to-catch use-after-free bugs) or the "release" build mode (optimized for run speed).
To run an entire test file row_cache_test, including all its test
functions, use a command like:
build/dev/test/boost/row_cache_test -- -c1 -m1G
to run a single test function test_reproduce_18045() from the longer test
file, use a command like:
build/dev/test/boost/row_cache_test -t test_reproduce_18045 -- -c1 -m1G
In these command lines, the parameters before the -- are passed to
Boost.Test, while the parameters after the -- are passed to the test code,
and in particular to Seastar. In this example Seastar is asked to run on one
CPU (-c1) and use 1G of memory (-m1G) instead of hogging the entire
machine. The Boost.Test option -t test_reproduce_18045 asks it to run just
this one test function instead of all the test functions in the executable.
Unfortunately, interrupting a running test with control-C while doesn't
work. This is a known bug (#5696). Kill a test with SIGKILL (-9) if you
need to kill it while it's running.
Boost tests can also be run using test.py - which is a script that provides a uniform way to run all tests in scylladb.git - C++ tests, Python tests, etc.
Writing tests
Because of the large build time and build size of each separate test executable, it is recommended to put test functions into relatively large source files. But not too large - to keep compilation time of a single source file (during development) at reasonable levels.
When adding new source files in test/boost, don't forget to list the new source file in configure.py and also in CMakeLists.txt. The former is needed by our CI, but the latter is preferred by some developers.