test(tablets): split and merge on writes and deletions

A test of high load, causing rapid tablet splits.
Then have a deletion burst and a major compaction to trigger rapid tablets merges.

jenkins-pipelines/oss/longevity/tablets-split-merge-test.jenkinsfile

@@ -0,0 +1,12 @@
+#!groovy
+
+// trick from https://github.com/jenkinsci/workflow-cps-global-lib-plugin/pull/43
+def lib = library identifier: 'sct@snapshot', retriever: legacySCM(scm)
+
+longevityPipeline(
+    backend: 'aws',
+    region: 'eu-west-1',
+    test_name: 'tablets_split_merge_test.TabletsSplitMergeTest.test_tablets_split_merge',
+    test_config: 'test-cases/features/tablets/tablets-split-merge-test.yaml'
+
+)

sdcm/utils/tablets/common.py

@@ -46,3 +46,19 @@ def wait_for_tablets_balanced(node):
                                 params={}, timeout=3600, retry=3)
         time.sleep(5)
     LOGGER.info("Tablets are balanced")
+
+
+def get_tablets_count(keyspace: str, session) -> int:
+    """
+    Returns the number of tablets in a keyspace.
+    :param session:
+    :param keyspace:
+    :return: tablets count int
+    """
+    cmd = f"select tablet_count,resize_type from system.tablets where keyspace_name='{keyspace}' ALLOW FILTERING"
+    cql_result = session.execute(cmd)
+    if not cql_result or not cql_result.current_rows:
+        return 0
+    tablet_count = cql_result.current_rows[0].tablet_count
+    LOGGER.debug('Retrieved a tablets count of: %s for keyspace %s', tablet_count, keyspace)
+    return int(tablet_count)

tablets_split_merge_test.py

@@ -0,0 +1,196 @@
+import math
+import re
+
+from cassandra import ConsistencyLevel
+from cassandra.query import SimpleStatement
+
+from longevity_test import LongevityTest
+from sdcm import wait
+from sdcm.sct_events import Severity
+from sdcm.sct_events.system import InfoEvent
+from sdcm.utils.tablets.common import get_tablets_count
+from test_lib.scylla_bench_tools import create_scylla_bench_table_query
+
+KEYSPACE_NAME = 'scylla_bench'
+TABLE_NAME = 'test'
+KS_CF = f"{KEYSPACE_NAME}.{TABLE_NAME}"
+
+
+class TabletsSplitMergeTest(LongevityTest):
+
+    def _get_tablets_number(self) -> int:
+        with self.db_cluster.cql_connection_patient(self.db_cluster.nodes[0]) as session:
+            tablets_number = get_tablets_count(keyspace=KEYSPACE_NAME, session=session)
+        self.log.debug(f"Tablets number for {KEYSPACE_NAME} is: {tablets_number}")
+        return tablets_number
+
+    def _get_max_expected_tablets(self, stress_cmd, target_tablet_size_in_bytes) -> int:
+        """
+        Example calculation:
+        partition_count = 500
+        clustering_row_count = 5555
+        clustering_row_size = 1884 bytes
+        target_tablet_size_in_bytes = 1073741824 bytes
+
+        Total Data Size=500×5555×1884=5,236,230,000 bytes≈5.2GB
+        Total Data Size=500×5555×1884=5,236,230,000bytes≈5.2GB
+
+        Now, dividing by the target tablet size of 1 GB:
+        Max Expected Tablets=5,236,230,000 / 1,073,741,824 ≈ 4.88 tablets
+        Thus, the system should expect 4 or 8 tablets based on this setup.
+
+        """
+        # Extract relevant parameters from stress_cmd
+        partition_count = int(re.search(r"-partition-count=(\d+)", stress_cmd).group(1))
+        clustering_row_count = int(re.search(r"-clustering-row-count=(\d+)", stress_cmd).group(1))
+        clustering_row_size = int(re.search(r"-clustering-row-size=(\d+)", stress_cmd).group(1))
+
+        # Calculate total data size (partition_count * clustering_row_count * clustering_row_size)
+        total_data_size = partition_count * clustering_row_count * clustering_row_size
+
+        # Calculate the maximum expected tablets by dividing total size by target tablet size
+        max_expected_tablets = total_data_size // target_tablet_size_in_bytes
+        self.log.debug(f"Calculated dataset size: {total_data_size / (1024 ** 3)} GB")
+        self.log.debug(f"Maximum expected tablets number: {max_expected_tablets}")
+        return max_expected_tablets
+
+    def _update_target_tablet_size(self, target_tablet_size_in_bytes: int):
+        for node in self.db_cluster.nodes:
+            self.log.info(f"Updating {node} with new target_tablet_size_in_bytes = {target_tablet_size_in_bytes}")
+            append_scylla_yaml = {"target_tablet_size_in_bytes": target_tablet_size_in_bytes}
+            with node.remote_scylla_yaml() as scylla_yaml:
+                scylla_yaml.update(append_scylla_yaml)
+            self.log.debug(f"Restarting node {node} to apply new target_tablet_size_in_bytes")
+            node.restart_scylla_server()
+
+    def _pre_create_large_partitions_schema(self):
+        self.run_pre_create_keyspace()
+        with self.db_cluster.cql_connection_patient(self.db_cluster.nodes[0]) as session:
+            session.execute(create_scylla_bench_table_query())
+
+    def test_tablets_split_merge(self):  # pylint: disable=too-many-locals  # noqa: PLR0914
+        """
+        (1) writing initial schema and get initial number of tablets.
+        (2) start write and read stress simultaneously.
+        (3) wait for tablets split.
+        (4) wait for stress completion.
+        (5) run deletions, nodetool flush and major compaction.
+        (6) wait for tablets merge, following a shrunk dataset size.
+        (7) redo more such cycles.
+        """
+
+        # (1) initiate schema.
+        InfoEvent(message=f"Create c-s keyspace with tablets initial value").publish()
+        self._pre_create_large_partitions_schema()
+        prepare_write_cmd = self.params.get('prepare_write_cmd')
+        stress_cmd = self.params.get('stress_cmd')
+        stress_read_cmd = self.params.get('stress_read_cmd')
+        append_scylla_yaml = self.params.get('append_scylla_yaml')
+        target_tablet_size_in_bytes = append_scylla_yaml['target_tablet_size_in_bytes']
+        max_expected_tablets = self._get_max_expected_tablets(
+            stress_cmd, target_tablet_size_in_bytes) + self._get_max_expected_tablets(prepare_write_cmd, target_tablet_size_in_bytes)
+        tablets_number_split_threshold = max_expected_tablets // 2
+        deletion_percentage = 70  # How many of dataset partitions should be deleted.
+        non_deleted_percentage = 100 - deletion_percentage
+        # In case, for example, 70% of data is deleted, then 30% left.
+        # Accordingly, it is expected to have about 30% of the tablets + some small extra (10%)
+        # example scenario:
+        # max_expected_tablets = 6
+        # deletion_percentage = 70, non_deleted_percentage = 30
+        # tablets_number_merge_threshold = math.ceil(6 * ((30+10) / 100) ) = 3
+        # Meaning there were 6 tablets, then after 70% of data is deleted, we expect 3 or fewer tablets.
+        tablets_number_merge_threshold = max_expected_tablets * math.ceil((non_deleted_percentage + 10) / 100)
+
+        # Run prepare stress
+        self.run_prepare_write_cmd()
+
+        # Run Read background stress
+        stress_read_queue = []
+        read_params = {'keyspace_num': 1, 'stress_cmd': stress_read_cmd}
+        self._run_all_stress_cmds(stress_read_queue, read_params)
+        self.node1 = self.db_cluster.nodes[0]
+
+        # Run cycles of writes and deletions
+        cycles_num = 3
+        for cycle in range(cycles_num):
+            tablets_num = self._get_tablets_number()
+            InfoEvent(message=f"Initial tablets number before stress is: {tablets_num}").publish()
+            InfoEvent(message=f"Starting write load: {stress_cmd}").publish()
+            stress_queue = []
+            self.assemble_and_run_all_stress_cmd(stress_queue=stress_queue, stress_cmd=stress_cmd, keyspace_num=1)
+
+            if self._is_stress_finished(stress_queue=stress_read_queue):
+                InfoEvent(message="The Read Stress is finished. Rerunning it.").publish()
+                self._run_all_stress_cmds(stress_read_queue, read_params)
+
+            InfoEvent(message="Wait for write stress to finish (if running)").publish()
+            for stress in stress_queue:
+                self.verify_stress_thread(cs_thread_pool=stress)
+
+            InfoEvent(message="Wait for tablets split following writes.").publish()
+            self._wait_for_more_tablet_than(tablets_number_split_threshold)
+
+            tablets_num = self._get_tablets_number()
+            InfoEvent(message=f"Start deletions to trigger a tablets merge following {tablets_num} tablets.").publish()
+            self.delete_partitions(deletion_percentage)
+            self.node1.run_nodetool(f"flush -- {KEYSPACE_NAME}")
+            self.node1.run_nodetool("compact", args=f"{KEYSPACE_NAME} {TABLE_NAME}")
+
+            InfoEvent(message="Wait for tablets merge after deletion is done.").publish()
+            self._wait_for_less_or_equal_tablet_than(tablets_number_merge_threshold)
+
+        InfoEvent(message="Wait for read stress to finish.").publish()
+
+        for stress in stress_read_queue:
+            self.verify_stress_thread(cs_thread_pool=stress)
+
+    def _is_stress_finished(self, stress_queue) -> bool:
+        for index, stress in enumerate(stress_queue):
+            self.log.debug(f"Checking stress task {index + 1}/{len(stress_queue)}: {stress.stress_cmd}")
+            if not all(future.done() for future in stress.results_futures):
+                self.log.debug(f"Stress task {index + 1} is still running.")
+                return False
+            self.log.debug(f"Stress task {index + 1} has completed.")
+        self.log.debug("All stress tasks have finished.")
+        return True
+
+    def _wait_for_less_or_equal_tablet_than(self, tablets_num: int, timeout_min: int = 15):
+        text = f"Waiting for a tablets number smaller than {tablets_num}"
+        res = wait.wait_for(func=lambda: self._get_tablets_number() <= tablets_num, step=60,
+                            text=text, timeout=60 * timeout_min, throw_exc=False)
+        if not res:
+            InfoEvent(message=f"{text} FAILED.", severity=Severity.ERROR).publish()
+
+    def _wait_for_more_tablet_than(self, tablets_num: int, timeout_min: int = 15):
+        text = f"Waiting for a tablets number bigger than {tablets_num}"
+        res = wait.wait_for(func=lambda: self._get_tablets_number() > tablets_num, step=60,
+                            text=text, timeout=60 * timeout_min, throw_exc=False)
+        if not res:
+            InfoEvent(message=f"{text} FAILED.", severity=Severity.ERROR).publish()
+
+    def delete_partitions(self, deletion_percentage: int):
+        """
+        Deletes a percentage of table's partitions (from the end of it).
+        Example:
+            partition_end_range = 10
+            max_partitions_to_delete = 100
+            deletion_percentage = 70
+            ==> 70 partitions from 11 up to 81 will be deleted.
+        """
+        start_partition = self.partitions_attrs.partition_end_range + 1
+        max_partitions_to_delete = self.partitions_attrs.max_partitions_in_test_table - self.partitions_attrs.partition_end_range
+        num_of_partitions_to_delete = int(max_partitions_to_delete * deletion_percentage / 100)
+        end_partition = start_partition + num_of_partitions_to_delete
+
+        self.log.debug(
+            f"Preparing to delete {num_of_partitions_to_delete} partitions from {start_partition} to {end_partition}.")
+        for pkey in range(start_partition, end_partition):
+            delete_query = f"delete from {KS_CF} where pk = {pkey}"
+            self.log.debug(f'delete query: {delete_query}')
+            try:
+                with self.db_cluster.cql_connection_patient(self.node1, connect_timeout=300) as session:
+                    session.execute(SimpleStatement(delete_query, consistency_level=ConsistencyLevel.QUORUM),
+                                    timeout=3600)
+            except Exception as e:  # pylint: disable=broad-except  # noqa: BLE001
+                message = f"Failed to execute delete: {e}"
+                InfoEvent(message=message, severity=Severity.ERROR).publish()

test-cases/features/tablets/tablets-split-merge-test.yaml

@@ -0,0 +1,26 @@
+test_duration: 240
+
+prepare_write_cmd: "scylla-bench -workload=sequential -mode=write -replication-factor=3 -partition-count=50 -clustering-row-count=5555                       -clustering-row-size=1884 -concurrency=20 -connection-count=20 -consistency-level=quorum -rows-per-request=10 -timeout=30s -validate-data"
+stress_cmd:        "scylla-bench -workload=sequential -mode=write -replication-factor=3 -partition-count=1000 -clustering-row-count=5555 -partition-offset=51 -clustering-row-size=1884 -concurrency=20 -connection-count=20 -consistency-level=quorum -rows-per-request=10 -timeout=30s -validate-data"
+stress_read_cmd:   "scylla-bench -workload=sequential -mode=read -replication-factor=3 -partition-count=50 -clustering-row-count=5555                        -clustering-row-size=1884 -concurrency=20 -connection-count=20 -consistency-level=quorum -rows-per-request=10 -timeout=30s -retry-number=30 -retry-interval=80ms,1s -iterations 0 -duration=30m -validate-data"
+
+pre_create_keyspace: "CREATE KEYSPACE IF NOT EXISTS scylla_bench WITH replication = {'class': 'NetworkTopologyStrategy', 'replication_factor': 3} and tablets={'initial':1};"
+
+data_validation: |
+  validate_partitions: true
+  table_name: "scylla_bench.test"
+  primary_key_column: "pk"
+  max_partitions_in_test_table: 1050
+  partition_range_with_data_validation: 0-50
+
+n_db_nodes: 3
+n_loaders: 1
+n_monitor_nodes: 1
+instance_type_db: 'i3.large'
+
+append_scylla_yaml:
+  target_tablet_size_in_bytes: 1073741824  # 1GB
+  enable_tablets: true
+
+user_prefix: 'tablets-split-merge'
+use_mgmt: false