use of org.apache.samza.checkpoint.Checkpoint in project samza by apache.
the class ContainerStorageManager method getBackendFactoryStoreNames.
/**
* Return a map of backend factory names to set of stores that should be restored using it
*/
@VisibleForTesting
Map<String, Set<String>> getBackendFactoryStoreNames(Checkpoint checkpoint, Set<String> storeNames, StorageConfig storageConfig) {
// backendFactoryName -> set(storeNames)
Map<String, Set<String>> backendFactoryStoreNames = new HashMap<>();
if (checkpoint != null && checkpoint.getVersion() == 1) {
// Only restore stores with changelog streams configured
Set<String> changelogStores = storeNames.stream().filter(storeName -> storageConfig.getChangelogStream(storeName).isPresent()).collect(Collectors.toSet());
// Default to changelog backend factory when using checkpoint v1 for backwards compatibility
if (!changelogStores.isEmpty()) {
backendFactoryStoreNames.put(StorageConfig.KAFKA_STATE_BACKEND_FACTORY, changelogStores);
}
if (storeNames.size() > changelogStores.size()) {
Set<String> nonChangelogStores = storeNames.stream().filter(storeName -> !changelogStores.contains(storeName)).collect(Collectors.toSet());
LOG.info("non-Side input stores: {}, do not have a configured store changelogs for checkpoint V1," + "restore for the store will be skipped", nonChangelogStores);
}
} else if (checkpoint == null || checkpoint.getVersion() == 2) {
// Extract the state checkpoint markers if checkpoint exists
Map<String, Map<String, String>> stateCheckpointMarkers = checkpoint == null ? Collections.emptyMap() : ((CheckpointV2) checkpoint).getStateCheckpointMarkers();
// Find stores associated to each state backend factory
storeNames.forEach(storeName -> {
List<String> storeFactories = storageConfig.getStoreRestoreFactories(storeName);
if (storeFactories.isEmpty()) {
// If the restore factory is not configured for the store and the store does not have a changelog topic
LOG.info("non-Side input store: {}, does not have a configured restore factories nor store changelogs," + "restore for the store will be skipped", storeName);
} else {
// Search the ordered list for the first matched state backend factory in the checkpoint
// If the checkpoint does not exist or state checkpoint markers does not exist, we match the first configured
// restore manager
Optional<String> factoryNameOpt = storeFactories.stream().filter(factoryName -> stateCheckpointMarkers.containsKey(factoryName) && stateCheckpointMarkers.get(factoryName).containsKey(storeName)).findFirst();
String factoryName;
if (factoryNameOpt.isPresent()) {
factoryName = factoryNameOpt.get();
} else {
// Restore factories configured but no checkpoints found
// Use first configured restore factory
factoryName = storeFactories.get(0);
LOG.warn("No matching checkpoints found for configured factories: {}, " + "defaulting to using the first configured factory with no checkpoints", storeFactories);
}
if (!backendFactoryStoreNames.containsKey(factoryName)) {
backendFactoryStoreNames.put(factoryName, new HashSet<>());
}
backendFactoryStoreNames.get(factoryName).add(storeName);
}
});
} else {
throw new SamzaException(String.format("Unsupported checkpoint version %s", checkpoint.getVersion()));
}
return backendFactoryStoreNames;
}
use of org.apache.samza.checkpoint.Checkpoint in project samza by apache.
the class TransactionalStateTaskRestoreManager method getStoreActions.
/**
* Marks each persistent but non-logged store for deletion.
*
* For each logged store, based on the current, checkpointed and local changelog offsets,
* 1. decides which directories (current and checkpoints) to delete for persistent stores.
* 2. decides which directories (checkpoints) to retain for persistent stores.
* 3. decides which stores (persistent or not) need to be restored, and the beginning and end offsets for the restore.
*
* When this method returns, in StoreActions,
* 1. all persistent store current directories will be present in storeDirsToDelete
* 2. each persistent store checkpoint directory will be present in either storeDirToRetain or storeDirsToDelete.
* 3. there will be at most one storeDirToRetain per persistent store, which will be a checkpoint directory.
* 4. any stores (persistent or not) that need to be restored from changelogs will be present in
* storesToRestore with appropriate offsets.
*/
@VisibleForTesting
static StoreActions getStoreActions(TaskModel taskModel, Map<String, StorageEngine> storeEngines, Map<String, SystemStream> storeChangelogs, Map<String, KafkaStateCheckpointMarker> kafkaStateCheckpointMarkers, CheckpointId checkpointId, Map<SystemStreamPartition, SystemStreamPartitionMetadata> currentChangelogOffsets, SystemAdmins systemAdmins, StorageManagerUtil storageManagerUtil, File loggedStoreBaseDirectory, File nonLoggedStoreBaseDirectory, Config config, Clock clock) {
TaskName taskName = taskModel.getTaskName();
TaskMode taskMode = taskModel.getTaskMode();
Map<String, File> storeDirToRetain = new HashMap<>();
ListMultimap<String, File> storeDirsToDelete = ArrayListMultimap.create();
Map<String, RestoreOffsets> storesToRestore = new HashMap<>();
storeEngines.forEach((storeName, storageEngine) -> {
// do nothing if store is non persistent and not logged (e.g. in memory cache only)
if (!storageEngine.getStoreProperties().isPersistedToDisk() && !storageEngine.getStoreProperties().isLoggedStore()) {
return;
}
// persistent but non-logged stores are always deleted
if (storageEngine.getStoreProperties().isPersistedToDisk() && !storageEngine.getStoreProperties().isLoggedStore()) {
File currentDir = storageManagerUtil.getTaskStoreDir(nonLoggedStoreBaseDirectory, storeName, taskName, taskMode);
LOG.info("Marking current directory: {} for store: {} in task: {} for deletion since it is not a logged store.", currentDir, storeName, taskName);
storeDirsToDelete.put(storeName, currentDir);
// persistent but non-logged stores should not have checkpoint dirs
return;
}
// get the oldest and newest current changelog SSP offsets as well as the checkpointed changelog SSP offset
SystemStream changelog = storeChangelogs.get(storeName);
SystemStreamPartition changelogSSP = new SystemStreamPartition(changelog, taskModel.getChangelogPartition());
SystemAdmin admin = systemAdmins.getSystemAdmin(changelogSSP.getSystem());
SystemStreamPartitionMetadata changelogSSPMetadata = currentChangelogOffsets.get(changelogSSP);
String oldestOffset = changelogSSPMetadata.getOldestOffset();
String newestOffset = changelogSSPMetadata.getNewestOffset();
// can be null if no message, or message has null offset
String checkpointedOffset = null;
if (kafkaStateCheckpointMarkers.containsKey(storeName) && StringUtils.isNotBlank(kafkaStateCheckpointMarkers.get(storeName).getChangelogOffset())) {
checkpointedOffset = kafkaStateCheckpointMarkers.get(storeName).getChangelogOffset();
}
long timeSinceLastCheckpointInMs = checkpointId == null ? Long.MAX_VALUE : System.currentTimeMillis() - checkpointId.getMillis();
// if the clean.store.start config is set, delete current and checkpoint dirs, restore from oldest offset to checkpointed
if (storageEngine.getStoreProperties().isPersistedToDisk() && new StorageConfig(config).cleanLoggedStoreDirsOnStart(storeName)) {
File currentDir = storageManagerUtil.getTaskStoreDir(loggedStoreBaseDirectory, storeName, taskName, taskMode);
LOG.info("Marking current directory: {} for store: {} in task: {} for deletion due to clean.on.container.start config.", currentDir, storeName, taskName);
storeDirsToDelete.put(storeName, currentDir);
storageManagerUtil.getTaskStoreCheckpointDirs(loggedStoreBaseDirectory, storeName, taskName, taskMode).forEach(checkpointDir -> {
LOG.info("Marking checkpoint directory: {} for store: {} in task: {} for deletion due to clean.on.container.start config.", checkpointDir, storeName, taskName);
storeDirsToDelete.put(storeName, checkpointDir);
});
LOG.info("Marking restore offsets for store: {} in task: {} to {}, {} ", storeName, taskName, oldestOffset, checkpointedOffset);
storesToRestore.put(storeName, new RestoreOffsets(oldestOffset, checkpointedOffset));
return;
}
Optional<File> currentDirOptional;
Optional<List<File>> checkpointDirsOptional;
if (!storageEngine.getStoreProperties().isPersistedToDisk()) {
currentDirOptional = Optional.empty();
checkpointDirsOptional = Optional.empty();
} else {
currentDirOptional = Optional.of(storageManagerUtil.getTaskStoreDir(loggedStoreBaseDirectory, storeName, taskName, taskMode));
checkpointDirsOptional = Optional.of(storageManagerUtil.getTaskStoreCheckpointDirs(loggedStoreBaseDirectory, storeName, taskName, taskMode));
}
LOG.info("For store: {} in task: {} got current dir: {}, checkpoint dirs: {}, checkpointed changelog offset: {}", storeName, taskName, currentDirOptional, checkpointDirsOptional, checkpointedOffset);
currentDirOptional.ifPresent(currentDir -> {
LOG.info("Marking current directory: {} for store: {} in task: {} for deletion.", currentDir, storeName, taskName);
storeDirsToDelete.put(storeName, currentDir);
});
if (checkpointedOffset == null && oldestOffset != null) {
// this can mean that either this is the initial migration for this feature and there are no previously
// checkpointed changelog offsets, or that this is a new store or changelog topic after the initial migration.
// if this is the first time migration, it might be desirable to retain existing data.
// if this is new store or topic, it is possible that the container previously died after writing some data to
// the changelog but before a commit, so it is desirable to delete the store, not restore anything and
// trim the changelog
// since we can't tell the difference b/w the two scenarios by just looking at the store and changelogs,
// we'll request users to indicate whether to retain existing data using a config flag. this flag should only
// be set during migrations, and turned off after the first successful commit of the new container (i.e. next
// deploy). for simplicity, we'll always delete the local store, and restore from changelog if necessary.
// the former scenario should not be common. the recommended way to opt-in to the transactional state feature
// is to first upgrade to the latest samza version but keep the transactional state restore config off.
// this will create the store checkpoint directories and write the changelog offset to the checkpoint, but
// will not use them during restore. once this is done (i.e. at least one commit after upgrade), the
// transactional state restore feature can be turned on on subsequent deploys. this code path exists as a
// fail-safe against clearing changelogs in case users do not follow upgrade instructions and enable the
// feature directly.
checkpointDirsOptional.ifPresent(checkpointDirs -> checkpointDirs.forEach(checkpointDir -> {
LOG.info("Marking checkpoint directory: {} for store: {} in task: {} for deletion since checkpointed " + "offset is null and oldest offset: {} is not.", checkpointDir, storeName, taskName, oldestOffset);
storeDirsToDelete.put(storeName, checkpointDir);
}));
if (new TaskConfig(config).getTransactionalStateRetainExistingState()) {
// mark for restore from (oldest, newest) to recreate local state.
LOG.warn("Checkpointed offset for store: {} in task: {} is null. Since retain existing state is true, " + "local state will be fully restored from current changelog contents. " + "There is no transactional local state guarantee.", storeName, taskName);
storesToRestore.put(storeName, new RestoreOffsets(oldestOffset, newestOffset));
} else {
LOG.warn("Checkpointed offset for store: {} in task: {} is null. Since retain existing state is false, " + "any local state and changelog topic contents will be deleted.", storeName, taskName);
// mark for restore from (oldest, null) to trim entire changelog.
storesToRestore.put(storeName, new RestoreOffsets(oldestOffset, null));
}
} else if (// check if the checkpointed offset is out of range of current oldest and newest offsets
admin.offsetComparator(oldestOffset, checkpointedOffset) > 0 || admin.offsetComparator(checkpointedOffset, newestOffset) > 0) {
// checkpointed offset is out of range. this could mean that this is a TTL topic and the checkpointed
// offset was TTLd, or that the changelog topic was manually deleted and then recreated.
// we cannot guarantee transactional state for TTL stores, so delete everything and do a full restore
// for local store. if the topic was deleted and recreated, this will have the side effect of
// clearing the store as well.
LOG.warn("Checkpointed offset: {} for store: {} in task: {} is out of range of oldest: {} or newest: {} offset." + "Deleting existing store and fully restoring from changelog topic from oldest to newest offset. If the topic " + "has time-based retention, there is no transactional local state guarantees. If the topic was changed," + "local state will be cleaned up and fully restored to match the new topic contents.", checkpointedOffset, storeName, taskName, oldestOffset, newestOffset);
checkpointDirsOptional.ifPresent(checkpointDirs -> checkpointDirs.forEach(checkpointDir -> storeDirsToDelete.put(storeName, checkpointDir)));
storesToRestore.put(storeName, new RestoreOffsets(oldestOffset, newestOffset));
} else {
// happy path. checkpointed offset is in range of current oldest and newest offsets
if (!checkpointDirsOptional.isPresent()) {
// non-persistent logged store
LOG.info("Did not find any checkpoint directories for logged (maybe non-persistent) store: {}. Local state " + "will be fully restored from current changelog contents.", storeName);
storesToRestore.put(storeName, new RestoreOffsets(oldestOffset, checkpointedOffset));
} else {
// persistent logged store
String targetOffset;
// check checkpoint time against min.compaction.lag.ms. if older, restore from checkpointed offset to newest
// with no trim. be conservative. allow 10% safety margin to avoid deletions when the downtime is close
// to min.compaction.lag.ms
long minCompactionLagMs = new StorageConfig(config).getChangelogMinCompactionLagMs(storeName);
if (timeSinceLastCheckpointInMs > .9 * minCompactionLagMs) {
LOG.warn("Checkpointed offset for store: {} in task: {} is: {}. It is in range of oldest: {} and " + "newest: {} changelog offset. However, time since last checkpoint is: {}, which is greater than " + "0.9 * min.compaction.lag.ms: {} for the changelog topic. Since there is a chance that" + "the changelog topic has been compacted, restoring store to the end of the current changelog contents." + "There is no transactional local state guarantee.", storeName, taskName, checkpointedOffset, oldestOffset, newestOffset, timeSinceLastCheckpointInMs, minCompactionLagMs);
targetOffset = newestOffset;
} else {
targetOffset = checkpointedOffset;
}
// if there exists a valid store checkpoint directory with oldest offset <= local offset <= target offset,
// retain it and restore the delta. delete all other checkpoint directories for the store. if more than one such
// checkpoint directory exists, retain the one with the highest local offset and delete the rest.
boolean hasValidCheckpointDir = false;
for (File checkpointDir : checkpointDirsOptional.get()) {
if (storageManagerUtil.isLoggedStoreValid(storeName, checkpointDir, config, storeChangelogs, taskModel, clock, storeEngines)) {
String localOffset = storageManagerUtil.readOffsetFile(checkpointDir, Collections.singleton(changelogSSP), false).get(changelogSSP);
LOG.info("Read local offset: {} for store: {} checkpoint dir: {} in task: {}", localOffset, storeName, checkpointDir, taskName);
if (admin.offsetComparator(localOffset, oldestOffset) >= 0 && admin.offsetComparator(localOffset, targetOffset) <= 0 && (storesToRestore.get(storeName) == null || admin.offsetComparator(localOffset, storesToRestore.get(storeName).startingOffset) > 0)) {
hasValidCheckpointDir = true;
LOG.info("Temporarily marking checkpoint dir: {} for store: {} in task: {} for retention. " + "May be overridden later.", checkpointDir, storeName, taskName);
storeDirToRetain.put(storeName, checkpointDir);
// mark for restore even if local == checkpointed, so that the changelog gets trimmed.
LOG.info("Temporarily marking store: {} in task: {} for restore from beginning offset: {} to " + "ending offset: {}. May be overridden later", storeName, taskName, localOffset, targetOffset);
storesToRestore.put(storeName, new RestoreOffsets(localOffset, targetOffset));
}
}
}
// delete all non-retained checkpoint directories
for (File checkpointDir : checkpointDirsOptional.get()) {
if (storeDirToRetain.get(storeName) == null || !storeDirToRetain.get(storeName).equals(checkpointDir)) {
LOG.info("Marking checkpoint directory: {} for store: {} in task: {} for deletion since it is not " + "marked for retention.", checkpointDir, storeName, taskName);
storeDirsToDelete.put(storeName, checkpointDir);
}
}
// if the store had not valid checkpoint dirs to retain, restore from changelog
if (!hasValidCheckpointDir) {
storesToRestore.put(storeName, new RestoreOffsets(oldestOffset, targetOffset));
}
}
}
});
LOG.info("Store directories to be retained in Task: {} are: {}", taskName, storeDirToRetain);
LOG.info("Store directories to be deleted in Task: {} are: {}", taskName, storeDirsToDelete);
LOG.info("Stores to be restored in Task: {} are: {}", taskName, storesToRestore);
return new StoreActions(storeDirToRetain, storeDirsToDelete, storesToRestore);
}
use of org.apache.samza.checkpoint.Checkpoint in project samza by apache.
the class BlobStoreRestoreManager method deleteUnusedStoresFromBlobStore.
/**
* Deletes blob store contents for stores that were present in the last checkpoint but are either no longer
* present in job configs (removed by user since last deployment) or are no longer configured to be backed
* up using blob stores.
*
* This method blocks until all the necessary store contents and snapshot index blobs have been marked for deletion.
*/
@VisibleForTesting
static void deleteUnusedStoresFromBlobStore(String jobName, String jobId, String taskName, StorageConfig storageConfig, BlobStoreConfig blobStoreConfig, Map<String, Pair<String, SnapshotIndex>> initialStoreSnapshotIndexes, BlobStoreUtil blobStoreUtil, ExecutorService executor) {
List<String> storesToBackup = storageConfig.getStoresWithBackupFactory(BlobStoreStateBackendFactory.class.getName());
List<String> storesToRestore = storageConfig.getStoresWithRestoreFactory(BlobStoreStateBackendFactory.class.getName());
List<CompletionStage<Void>> storeDeletionFutures = new ArrayList<>();
initialStoreSnapshotIndexes.forEach((storeName, scmAndSnapshotIndex) -> {
if (!storesToBackup.contains(storeName) && !storesToRestore.contains(storeName)) {
LOG.debug("Removing task: {} store: {} from blob store. It is either no longer used, " + "or is no longer configured to be backed up or restored with blob store.", taskName, storeName);
DirIndex dirIndex = scmAndSnapshotIndex.getRight().getDirIndex();
Metadata requestMetadata = new Metadata(Metadata.SNAPSHOT_INDEX_PAYLOAD_PATH, Optional.empty(), jobName, jobId, taskName, storeName);
CompletionStage<Void> storeDeletionFuture = // delete files and sub-dirs previously marked for removal
blobStoreUtil.cleanUpDir(dirIndex, requestMetadata).thenComposeAsync(v -> blobStoreUtil.deleteDir(dirIndex, requestMetadata), // deleted files and dirs still present
executor).thenComposeAsync(v -> blobStoreUtil.deleteSnapshotIndexBlob(scmAndSnapshotIndex.getLeft(), requestMetadata), // delete the snapshot index blob
executor);
storeDeletionFutures.add(storeDeletionFuture);
}
});
FutureUtil.allOf(storeDeletionFutures).join();
}
use of org.apache.samza.checkpoint.Checkpoint in project samza by apache.
the class TaskStorageCommitManager method init.
public void init() {
// Assuming that container storage manager has already started and created to stores
storageEngines = containerStorageManager.getAllStores(taskName);
if (checkpointManager != null) {
Checkpoint checkpoint = checkpointManager.readLastCheckpoint(taskName);
LOG.debug("Last checkpoint on start for task: {} is: {}", taskName, checkpoint);
stateBackendToBackupManager.values().forEach(storageBackupManager -> storageBackupManager.init(checkpoint));
} else {
stateBackendToBackupManager.values().forEach(storageBackupManager -> storageBackupManager.init(null));
}
}
use of org.apache.samza.checkpoint.Checkpoint in project samza by apache.
the class TestBlobStoreUtil method testGetSSISkipsStoresWithSnapshotIndexAlreadyDeleted.
@Test
public void testGetSSISkipsStoresWithSnapshotIndexAlreadyDeleted() {
String store = "storeName1";
String otherStore = "storeName2";
Checkpoint checkpoint = createCheckpointV2(BlobStoreStateBackendFactory.class.getName(), ImmutableMap.of(store, "snapshotIndexBlobId1", otherStore, "snapshotIndexBlobId2"));
Set<String> storesToBackupOrRestore = new HashSet<>();
storesToBackupOrRestore.add(store);
storesToBackupOrRestore.add(otherStore);
SnapshotIndex store1SnapshotIndex = mock(SnapshotIndex.class);
BlobStoreUtil mockBlobStoreUtil = mock(BlobStoreUtil.class);
CompletableFuture<SnapshotIndex> failedFuture = FutureUtil.failedFuture(new DeletedException());
when(mockBlobStoreUtil.getSnapshotIndex(eq("snapshotIndexBlobId1"), any(Metadata.class))).thenReturn(CompletableFuture.completedFuture(store1SnapshotIndex));
when(mockBlobStoreUtil.getSnapshotIndex(eq("snapshotIndexBlobId2"), any(Metadata.class))).thenReturn(failedFuture);
when(mockBlobStoreUtil.getStoreSnapshotIndexes(anyString(), anyString(), anyString(), any(Checkpoint.class), anySetOf(String.class))).thenCallRealMethod();
Map<String, Pair<String, SnapshotIndex>> snapshotIndexes = mockBlobStoreUtil.getStoreSnapshotIndexes("testJobName", "testJobId", "taskName", checkpoint, storesToBackupOrRestore);
assertEquals(1, snapshotIndexes.size());
assertEquals("snapshotIndexBlobId1", snapshotIndexes.get("storeName1").getLeft());
assertEquals(store1SnapshotIndex, snapshotIndexes.get("storeName1").getRight());
}
Aggregations