@powersync/service-module-mongodb 0.16.0 → 0.18.0

package/dist/replication/ChangeStream.js

@@ -7,7 +7,7 @@ import { MongoLSN } from '../common/MongoLSN.js';
 import { PostImagesOption } from '../types/types.js';
 import { escapeRegExp } from '../utils.js';
 import { createCheckpoint, getCacheIdentifier, getMongoRelation, STANDALONE_CHECKPOINT_ID } from './MongoRelation.js';
-import { ChunkedSnapshotQuery } from './MongoSnapshotQuery.js';
+import { MongoSnapshotter } from './MongoSnapshotter.js';
 import { parseChangeDocument, rawChangeStream } from './RawChangeStream.js';
 import { CHECKPOINTS_COLLECTION, timestampToDate } from './replication-utils.js';
 import { DirectSourceRowConverter } from './SourceRowConverter.js';
@@ -34,21 +34,29 @@ export class ChangeStream {
     defaultDb;
     metrics;
     maxAwaitTimeMS;
-    abort_signal;
+    abortController = new AbortController();
+    abortSignal = this.abortController.signal;
+    initPromise = null;
+    snapshotter;
+    /**
+     * We use the relationCache _only_ for caching static SourceTable info, not for snapshot status.
+     */
     relationCache = new RelationCache(getCacheIdentifier);
     replicationLag = new ReplicationLagTracker();
     checkpointStreamId = new mongo.ObjectId();
     logger;
     snapshotChunkLength;
     changeStreamTimeout;
+    storageHooks;
     sourceRowConverter;
     constructor(options) {
         this.storage = options.storage;
         this.metrics = options.metrics;
-        this.group_id = options.storage.group_id;
+        this.group_id = options.storage.replicationStreamId;
         this.connections = options.connections;
         this.maxAwaitTimeMS = options.maxAwaitTimeMS ?? 10_000;
         this.snapshotChunkLength = options.snapshotChunkLength ?? 6_000;
+        this.storageHooks = options.storageHooks;
         this.client = this.connections.client;
         this.defaultDb = this.connections.db;
         this.sync_rules = options.storage.getParsedSyncRules({
@@ -58,14 +66,26 @@ export class ChangeStream {
         // The change stream aggregation command should timeout before the socket times out,
         // so we use 90% of the socket timeout value.
         this.changeStreamTimeout = Math.ceil(this.client.options.socketTimeoutMS * 0.9);
-        this.abort_signal = options.abort_signal;
-        this.abort_signal.addEventListener('abort', () => {
-            // TODO: Fast abort?
+        const baseLogger = options.logger ?? this.storage.logger;
+        // Unfortunately the Winston APIs don't have a nice way to append to the prefix,
+        // so we replace it here.
+        this.logger = baseLogger.child({ prefix: `[${this.storage.replicationStreamName}] [stream] ` });
+        const snapshotLogger = baseLogger.child({ prefix: `[${this.storage.replicationStreamName}] [snapshot] ` });
+        this.snapshotter = new MongoSnapshotter({
+            ...options,
+            abortSignal: this.abortSignal,
+            logger: snapshotLogger,
+            checkpointStreamId: this.checkpointStreamId
+        });
+        options.abort_signal.addEventListener('abort', () => {
+            this.abortController.abort(options.abort_signal.reason);
         }, { once: true });
-        this.logger = options.logger ?? this.storage.logger;
+        if (options.abort_signal.aborted) {
+            this.abortController.abort(options.abort_signal.reason);
+        }
     }
     get stopped() {
-        return this.abort_signal.aborted;
+        return this.abortSignal.aborted;
     }
     get usePostImages() {
         return this.connections.options.postImages != PostImagesOption.OFF;
@@ -73,229 +93,6 @@ export class ChangeStream {
     get configurePostImages() {
         return this.connections.options.postImages == PostImagesOption.AUTO_CONFIGURE;
     }
-    /**
-     * This resolves a pattern, persists the related metadata, and returns
-     * the resulting SourceTables.
-     *
-     * This implicitly checks the collection postImage configuration.
-     */
-    async resolveQualifiedTableNames(batch, tablePattern) {
-        const schema = tablePattern.schema;
-        if (tablePattern.connectionTag != this.connections.connectionTag) {
-            return [];
-        }
-        let nameFilter;
-        if (tablePattern.isWildcard) {
-            nameFilter = new RegExp('^' + escapeRegExp(tablePattern.tablePrefix));
-        }
-        else {
-            nameFilter = tablePattern.name;
-        }
-        let result = [];
-        // Check if the collection exists
-        const collections = await this.client
-            .db(schema)
-            .listCollections({
-            name: nameFilter
-        }, { nameOnly: false })
-            .toArray();
-        if (!tablePattern.isWildcard && collections.length == 0) {
-            this.logger.warn(`Collection ${schema}.${tablePattern.name} not found`);
-        }
-        for (let collection of collections) {
-            const table = await this.handleRelation(batch, getMongoRelation({ db: schema, coll: collection.name }), 
-            // This is done as part of the initial setup - snapshot is handled elsewhere
-            { snapshot: false, collectionInfo: collection });
-            result.push(table);
-        }
-        return result;
-    }
-    async initSlot() {
-        const status = await this.storage.getStatus();
-        if (status.snapshot_done && status.checkpoint_lsn) {
-            this.logger.info(`Initial replication already done`);
-            return { needsInitialSync: false, snapshotLsn: null };
-        }
-        return { needsInitialSync: true, snapshotLsn: status.snapshot_lsn };
-    }
-    async estimatedCount(table) {
-        const count = await this.estimatedCountNumber(table);
-        return `~${count}`;
-    }
-    async estimatedCountNumber(table) {
-        const db = this.client.db(table.schema);
-        return await db.collection(table.name).estimatedDocumentCount();
-    }
-    /**
-     * This gets a LSN before starting a snapshot, which we can resume streaming from after the snapshot.
-     *
-     * This LSN can survive initial replication restarts.
-     */
-    async getSnapshotLsn() {
-        const hello = await this.defaultDb.command({ hello: 1 });
-        // Basic sanity check
-        if (hello.msg == 'isdbgrid') {
-            throw new ServiceError(ErrorCode.PSYNC_S1341, 'Sharded MongoDB Clusters are not supported yet (including MongoDB Serverless instances).');
-        }
-        else if (hello.setName == null) {
-            throw new ServiceError(ErrorCode.PSYNC_S1342, 'Standalone MongoDB instances are not supported - use a replicaset.');
-        }
-        // Open a change stream just to get a resume token for later use.
-        // We could use clusterTime from the hello command, but that won't tell us if the
-        // snapshot isn't valid anymore.
-        // If we just use the first resumeToken from the stream, we get two potential issues:
-        // 1. The resumeToken may just be a wrapped clusterTime, which does not detect changes
-        //    in source db or other stream issues.
-        // 2. The first actual change we get may have the same clusterTime, causing us to incorrect
-        //    skip that event.
-        // Instead, we create a new checkpoint document, and wait until we get that document back in the stream.
-        // To avoid potential race conditions with the checkpoint creation, we create a new checkpoint document
-        // periodically until the timeout is reached.
-        const LSN_TIMEOUT_SECONDS = 60;
-        const LSN_CREATE_INTERVAL_SECONDS = 1;
-        // Create a checkpoint, and open a change stream using startAtOperationTime with the checkpoint's operationTime.
-        const firstCheckpointLsn = await createCheckpoint(this.client, this.defaultDb, this.checkpointStreamId);
-        const startTime = performance.now();
-        let lastCheckpointCreated = performance.now();
-        let eventsSeen = 0;
-        let batchesSeen = 0;
-        const filters = this.getSourceNamespaceFilters();
-        const iter = this.rawChangeStreamBatches({
-            lsn: firstCheckpointLsn,
-            maxAwaitTimeMS: 0,
-            signal: this.abort_signal,
-            filters
-        });
-        for await (let { events } of iter) {
-            if (performance.now() - startTime >= LSN_TIMEOUT_SECONDS * 1000) {
-                break;
-            }
-            if (performance.now() - lastCheckpointCreated >= LSN_CREATE_INTERVAL_SECONDS * 1000) {
-                await createCheckpoint(this.client, this.defaultDb, this.checkpointStreamId);
-                lastCheckpointCreated = performance.now();
-            }
-            batchesSeen += 1;
-            for (let rawChangeDocument of events) {
-                const changeDocument = parseChangeDocument(rawChangeDocument);
-                const ns = 'ns' in changeDocument && 'coll' in changeDocument.ns ? changeDocument.ns : undefined;
-                if (ns?.coll == CHECKPOINTS_COLLECTION && 'documentKey' in changeDocument) {
-                    const checkpointId = changeDocument.documentKey._id;
-                    if (!this.checkpointStreamId.equals(checkpointId)) {
-                        continue;
-                    }
-                    const { comparable: lsn } = new MongoLSN({
-                        timestamp: changeDocument.clusterTime,
-                        resume_token: changeDocument._id
-                    });
-                    return lsn;
-                }
-                eventsSeen += 1;
-            }
-        }
-        // Could happen if there is a very large replication lag?
-        throw new ServiceError(ErrorCode.PSYNC_S1301, `Timeout after while waiting for checkpoint document for ${LSN_TIMEOUT_SECONDS}s. Streamed events = ${eventsSeen}, batches = ${batchesSeen}`);
-    }
-    /**
-     * Given a snapshot LSN, validate that we can read from it, by opening a change stream.
-     */
-    async validateSnapshotLsn(lsn) {
-        const filters = this.getSourceNamespaceFilters();
-        const stream = this.rawChangeStreamBatches({
-            lsn: lsn,
-            // maxAwaitTimeMS should never actually be used here
-            maxAwaitTimeMS: 0,
-            filters
-        });
-        for await (let _batch of stream) {
-            // We got a response from the aggregate command, so consider the LSN valid.
-            // Close the stream immediately.
-            break;
-        }
-    }
-    async initialReplication(snapshotLsn) {
-        const sourceTables = this.sync_rules.getSourceTables();
-        await this.client.connect();
-        const tracer = new PerformanceTracer('MongoDB initial replication');
-        const flushResult = await this.storage.startBatch({
-            logger: this.logger,
-            zeroLSN: MongoLSN.ZERO.comparable,
-            defaultSchema: this.defaultDb.databaseName,
-            storeCurrentData: false,
-            skipExistingRows: true,
-            tracer
-        }, async (batch) => {
-            if (snapshotLsn == null) {
-                // First replication attempt - get a snapshot and store the timestamp
-                snapshotLsn = await this.getSnapshotLsn();
-                await batch.setResumeLsn(snapshotLsn);
-                this.logger.info(`Marking snapshot at ${snapshotLsn}`);
-            }
-            else {
-                this.logger.info(`Resuming snapshot at ${snapshotLsn}`);
-                // Check that the snapshot is still valid.
-                await this.validateSnapshotLsn(snapshotLsn);
-            }
-            // Start by resolving all tables.
-            // This checks postImage configuration, and that should fail as
-            // early as possible.
-            let allSourceTables = [];
-            for (let tablePattern of sourceTables) {
-                const tables = await this.resolveQualifiedTableNames(batch, tablePattern);
-                allSourceTables.push(...tables);
-            }
-            let tablesWithStatus = [];
-            for (let table of allSourceTables) {
-                if (table.snapshotComplete) {
-                    this.logger.info(`Skipping ${table.qualifiedName} - snapshot already done`);
-                    continue;
-                }
-                let count = await this.estimatedCountNumber(table);
-                const updated = await batch.updateTableProgress(table, {
-                    totalEstimatedCount: count
-                });
-                tablesWithStatus.push(updated);
-                this.relationCache.update(updated);
-                this.logger.info(`To replicate: ${table.qualifiedName}: ${updated.snapshotStatus?.replicatedCount}/~${updated.snapshotStatus?.totalEstimatedCount}`);
-            }
-            for (let table of tablesWithStatus) {
-                await this.snapshotTable(batch, table);
-                await batch.markTableSnapshotDone([table]);
-                this.touch();
-            }
-            // The checkpoint here is a marker - we need to replicate up to at least this
-            // point before the data can be considered consistent.
-            // We could do this for each individual table, but may as well just do it once for the entire snapshot.
-            const checkpoint = await createCheckpoint(this.client, this.defaultDb, STANDALONE_CHECKPOINT_ID);
-            await batch.markAllSnapshotDone(checkpoint);
-            // This will not create a consistent checkpoint yet, but will persist the op.
-            // Actual checkpoint will be created when streaming replication caught up.
-            await batch.commit(snapshotLsn);
-            this.logger.info(`Snapshot done. Need to replicate from ${snapshotLsn} to ${checkpoint} to be consistent`);
-        });
-        return { lastOpId: flushResult?.flushed_op };
-    }
-    async setupCheckpointsCollection() {
-        const collection = await this.getCollectionInfo(this.defaultDb.databaseName, CHECKPOINTS_COLLECTION);
-        if (collection == null) {
-            await this.defaultDb.createCollection(CHECKPOINTS_COLLECTION, {
-                changeStreamPreAndPostImages: { enabled: true }
-            });
-        }
-        else if (this.usePostImages && collection.options?.changeStreamPreAndPostImages?.enabled != true) {
-            // Drop + create requires less permissions than collMod,
-            // and we don't care about the data in this collection.
-            await this.defaultDb.dropCollection(CHECKPOINTS_COLLECTION);
-            await this.defaultDb.createCollection(CHECKPOINTS_COLLECTION, {
-                changeStreamPreAndPostImages: { enabled: true }
-            });
-        }
-        else {
-            // Clear the collection on startup, to keep it clean
-            // We never query this collection directly, and don't want to keep the data around.
-            // We only use this to get data into the oplog/changestream.
-            await this.defaultDb.collection(CHECKPOINTS_COLLECTION).deleteMany({});
-        }
-    }
     getSourceNamespaceFilters() {
         const sourceTables = this.sync_rules.getSourceTables();
         let $inFilters = [
@@ -343,73 +140,10 @@ export class ChangeStream {
         }
         return { $match: nsFilter, multipleDatabases };
     }
-    async snapshotTable(batch, table) {
-        const rowsReplicatedMetric = this.metrics.getCounter(ReplicationMetric.ROWS_REPLICATED);
-        const bytesReplicatedMetric = this.metrics.getCounter(ReplicationMetric.DATA_REPLICATED_BYTES);
-        const chunksReplicatedMetric = this.metrics.getCounter(ReplicationMetric.CHUNKS_REPLICATED);
-        const totalEstimatedCount = await this.estimatedCountNumber(table);
-        let at = table.snapshotStatus?.replicatedCount ?? 0;
-        const db = this.client.db(table.schema);
-        const collection = db.collection(table.name);
-        await using query = new ChunkedSnapshotQuery({
-            collection,
-            key: table.snapshotStatus?.lastKey,
-            batchSize: this.snapshotChunkLength
-        });
-        if (query.lastKey != null) {
-            this.logger.info(`Replicating ${table.qualifiedName} ${table.formatSnapshotProgress()} - resuming at _id > ${query.lastKey}`);
-        }
-        else {
-            this.logger.info(`Replicating ${table.qualifiedName} ${table.formatSnapshotProgress()}`);
-        }
-        let lastBatch = performance.now();
-        let nextChunkPromise = query.nextChunk();
-        while (true) {
-            const { docs: docBatch, lastKey, bytes: chunkBytes } = await nextChunkPromise;
-            if (docBatch.length == 0) {
-                // No more data - stop iterating
-                break;
-            }
-            bytesReplicatedMetric.add(chunkBytes);
-            chunksReplicatedMetric.add(1);
-            if (this.abort_signal.aborted) {
-                throw new ReplicationAbortedError(`Aborted initial replication`, this.abort_signal.reason);
-            }
-            // Pre-fetch next batch, so that we can read and write concurrently
-            nextChunkPromise = query.nextChunk();
-            for (let buffer of docBatch) {
-                const { row: record, replicaId: replicaId } = this.rawToSqliteRow(buffer);
-                // This auto-flushes when the batch reaches its size limit
-                await batch.save({
-                    tag: SaveOperationTag.INSERT,
-                    sourceTable: table,
-                    before: undefined,
-                    beforeReplicaId: undefined,
-                    after: record,
-                    afterReplicaId: replicaId
-                });
-            }
-            // Important: flush before marking progress
-            await batch.flush();
-            at += docBatch.length;
-            rowsReplicatedMetric.add(docBatch.length);
-            table = await batch.updateTableProgress(table, {
-                lastKey,
-                replicatedCount: at,
-                totalEstimatedCount: totalEstimatedCount
-            });
-            this.relationCache.update(table);
-            const duration = performance.now() - lastBatch;
-            lastBatch = performance.now();
-            this.logger.info(`Replicating ${table.qualifiedName} ${table.formatSnapshotProgress()} in ${duration.toFixed(0)}ms`);
-            this.touch();
-        }
-        // In case the loop was interrupted, make sure we await the last promise.
-        await nextChunkPromise;
-    }
-    async getRelation(batch, descriptor, options) {
-        const existing = this.relationCache.get(descriptor);
+    async getRelations(batch, descriptor, options) {
+        const existing = this.relationCache.getAll(descriptor);
         if (existing != null) {
+            // We do this even when it's an empty result: Empty means nothing to sync, and we don't need to re-resolve.
             return existing;
         }
         // Note: collection may have been dropped at this point, so we handle
@@ -452,14 +186,11 @@ export class ChangeStream {
             // Ignore the postImages check in this case.
         }
         const snapshot = options.snapshot;
-        const result = await this.storage.resolveTable({
-            group_id: this.group_id,
+        const result = await batch.resolveTables({
             connection_id: this.connection_id,
-            connection_tag: this.connections.connectionTag,
-            entity_descriptor: descriptor,
-            sync_rules: this.sync_rules
+            source: descriptor
         });
-        this.relationCache.update(result.table);
+        this.relationCache.updateAll(descriptor, result.tables);
         // Drop conflicting collections.
         // This is generally not expected for MongoDB source dbs, so we log an error.
         if (result.dropTables.length > 0) {
@@ -470,17 +201,12 @@ export class ChangeStream {
         // 1. Snapshot is requested (false for initial snapshot, since that process handles it elsewhere)
         // 2. Snapshot is not already done, AND:
         // 3. The table is used in sync config.
-        const shouldSnapshot = snapshot && !result.table.snapshotComplete && result.table.syncAny;
-        if (shouldSnapshot) {
+        const snapshotCandidates = result.tables.filter((table) => snapshot && !table.snapshotComplete && table.syncAny);
+        if (snapshotCandidates.length > 0) {
             this.logger.info(`New collection: ${descriptor.schema}.${descriptor.name}`);
-            // Truncate this table, in case a previous snapshot was interrupted.
-            await batch.truncate([result.table]);
-            await this.snapshotTable(batch, result.table);
-            const no_checkpoint_before_lsn = await createCheckpoint(this.client, this.defaultDb, STANDALONE_CHECKPOINT_ID);
-            const [table] = await batch.markTableSnapshotDone([result.table], no_checkpoint_before_lsn);
-            return table;
-        }
-        return result.table;
+            await this.snapshotter.snapshotTables(batch, snapshotCandidates);
+        }
+        return result.tables;
     }
     async writeChange(batch, table, change) {
         if (!table.syncAny) {
@@ -535,36 +261,76 @@ export class ChangeStream {
         }
     }
     async replicate() {
+        let streamPromise = null;
+        let loopPromise = null;
+        let cleanupPromise = null;
         try {
             // If anything errors here, the entire replication process is halted, and
             // all connections automatically closed, including this one.
-            await this.initReplication();
-            await this.streamChanges();
+            this.initPromise = this.initReplication();
+            await this.initPromise;
+            loopPromise = this.snapshotter
+                .replicationLoop()
+                .then(() => {
+                throw new ReplicationAssertionError(`Replication snapshotter exited unexpectedly`);
+            })
+                .catch((e) => {
+                this.abortController.abort(e);
+                throw e;
+            });
+            if (!this.snapshotter.supportsConcurrentSnapshots) {
+                await Promise.race([this.snapshotter.waitForInitialSnapshot(), loopPromise]);
+            }
+            // Unlike the other two, this resolves on completion, not an indefinite loop.
+            cleanupPromise = this.cleanupStoppedSyncConfigs().catch((e) => {
+                this.abortController.abort(e);
+                throw e;
+            });
+            streamPromise = this.streamChanges()
+                .then(() => {
+                throw new ReplicationAssertionError(`Replication stream exited unexpectedly`);
+            })
+                .catch((e) => {
+                this.abortController.abort(e);
+                throw e;
+            });
+            const results = await Promise.allSettled([loopPromise, streamPromise, cleanupPromise]);
+            throw replicationLoopError(results);
         }
         catch (e) {
             await this.storage.reportError(e);
             throw e;
         }
+        finally {
+            this.abortController.abort();
+        }
+    }
+    async waitForInitialSnapshot() {
+        if (this.initPromise == null) {
+            throw new ReplicationAssertionError('replicate() must be called before waitForInitialSnapshot()');
+        }
+        await this.initPromise;
+        await this.snapshotter.waitForInitialSnapshot();
     }
     async initReplication() {
-        const result = await this.initSlot();
-        await this.setupCheckpointsCollection();
+        const result = await this.snapshotter.checkSlot();
+        await this.snapshotter.setupCheckpointsCollection();
         if (result.needsInitialSync) {
             if (result.snapshotLsn == null) {
                 // Snapshot LSN is not present, so we need to start replication from scratch.
-                await this.storage.clear({ signal: this.abort_signal });
-            }
-            const { lastOpId } = await this.initialReplication(result.snapshotLsn);
-            if (lastOpId != null) {
-                // Populate the cache _after_ initial replication, but _before_ we switch to this replication stream.
-                await this.storage.populatePersistentChecksumCache({
-                    signal: this.abort_signal,
-                    // No checkpoint yet, but we do have the opId.
-                    maxOpId: lastOpId
-                });
+                await this.storage.clear({ signal: this.abortSignal });
             }
+            await this.snapshotter.queueSnapshotTables(result.snapshotLsn);
         }
     }
+    async cleanupStoppedSyncConfigs() {
+        await this.storage.cleanupStoppedSyncConfigs?.({
+            signal: this.abortSignal,
+            logger: this.logger,
+            defaultSchema: this.defaultDb.databaseName,
+            sourceConnectionTag: this.connections.connectionTag
+        });
+    }
     async streamChanges() {
         try {
             await this.streamChangesInternal();
@@ -649,6 +415,7 @@ export class ChangeStream {
             defaultSchema: this.defaultDb.databaseName,
             // We get a complete postimage for every change, so we don't need to store the current data.
             storeCurrentData: false,
+            hooks: this.storageHooks,
             tracer
         }, async (batch) => {
             const { resumeFromLsn } = batch;
@@ -667,7 +434,7 @@ export class ChangeStream {
             const batchStream = this.rawChangeStreamBatches({
                 lsn: resumeFromLsn,
                 filters,
-                signal: this.abort_signal,
+                signal: this.abortSignal,
                 tracer
             });
             // Always start with a checkpoint.
@@ -683,7 +450,7 @@ export class ChangeStream {
                 using batchSpan = tracer.span('processing');
                 bytesReplicatedMetric.add(eventBatch.byteSize);
                 chunksReplicatedMetric.add(1);
-                if (this.abort_signal.aborted) {
+                if (this.abortSignal.aborted) {
                     break;
                 }
                 this.touch();
@@ -712,7 +479,7 @@ export class ChangeStream {
                 for (let eventIndex = 0; eventIndex < events.length; eventIndex++) {
                     const rawChangeDocument = events[eventIndex];
                     const originalChangeDocument = parseChangeDocument(rawChangeDocument);
-                    if (this.abort_signal.aborted) {
+                    if (this.abortSignal.aborted) {
                         break;
                     }
                     if (startAfter != null && originalChangeDocument.clusterTime?.lte(startAfter)) {
@@ -799,12 +566,19 @@ export class ChangeStream {
                             // change stream events, collapse standalone checkpoints into the normal batch
                             // checkpoint flow to avoid commit churn under sustained load.
                             const hasBufferedChanges = eventIndex < events.length - 1;
-                            if (waitForCheckpointLsn != null || hasBufferedChanges) {
-                                if (waitForCheckpointLsn == null) {
-                                    waitForCheckpointLsn = await createCheckpoint(this.client, this.defaultDb, this.checkpointStreamId);
-                                }
+                            if (hasBufferedChanges && waitForCheckpointLsn == null) {
+                                // Buffered changes - create a new batch checkpoint to rate limit commits
+                                using _ = tracer.span('source_checkpoint');
+                                waitForCheckpointLsn = await createCheckpoint(this.client, this.defaultDb, this.checkpointStreamId);
                                 continue;
                             }
+                            else if (waitForCheckpointLsn != null) {
+                                // Skip this checkpoint - wait for the batch checkpoint.
+                                continue;
+                            }
+                            else {
+                                // No buffered changes, and no batch checkpoint pending - commit immediately.
+                            }
                         }
                         else if (!this.checkpointStreamId.equals(checkpointId)) {
                             continue;
@@ -813,21 +587,13 @@ export class ChangeStream {
                             timestamp: changeDocument.clusterTime,
                             resume_token: changeDocument._id
                         });
-                        if (batch.lastCheckpointLsn != null && lsn < batch.lastCheckpointLsn) {
-                            // Checkpoint out of order - should never happen with MongoDB.
-                            // If it does happen, we throw an error to stop the replication - restarting should recover.
-                            // Since we use batch.lastCheckpointLsn for the next resumeAfter, this should not result in an infinite loop.
-                            // Originally a workaround for https://jira.mongodb.org/browse/NODE-7042.
-                            // This has been fixed in the driver in the meantime, but we still keep this as a safety-check.
-                            throw new ReplicationAssertionError(`Change resumeToken ${changeDocument._id._data} (${timestampToDate(changeDocument.clusterTime).toISOString()}) is less than last checkpoint LSN ${batch.lastCheckpointLsn}. Restarting replication.`);
-                        }
                         if (waitForCheckpointLsn != null && lsn >= waitForCheckpointLsn) {
                             waitForCheckpointLsn = null;
                         }
-                        const { checkpointBlocked } = await batch.commit(lsn, {
+                        const { checkpointBlocked, checkpointCreated } = await batch.commit(lsn, {
                             oldestUncommittedChange: this.replicationLag.oldestUncommittedChange
                         });
-                        if (!checkpointBlocked) {
+                        if (!checkpointBlocked || checkpointCreated) {
                             this.replicationLag.markCommitted();
                         }
                     }
@@ -836,17 +602,19 @@ export class ChangeStream {
                         changeDocument.operationType == 'replace' ||
                         changeDocument.operationType == 'delete') {
                         if (waitForCheckpointLsn == null) {
+                            using _ = tracer.span('source_checkpoint');
                             waitForCheckpointLsn = await createCheckpoint(this.client, this.defaultDb, this.checkpointStreamId);
                         }
-                        const rel = getMongoRelation(changeDocument.ns);
-                        const table = await this.getRelation(batch, rel, {
+                        const rel = getMongoRelation(changeDocument.ns, this.connections.connectionTag);
+                        const tables = await this.getRelations(batch, rel, {
                             // In most cases, we should not need to snapshot this. But if this is the first time we see the collection
                             // for whatever reason, then we do need to snapshot it.
                             // This may result in some duplicate operations when a collection is created for the first time after
                             // sync config was deployed.
                             snapshot: true
                         });
-                        if (table.syncAny) {
+                        const tablesToReplicate = tables.filter((table) => table.syncAny);
+                        if (tablesToReplicate.length > 0) {
                             this.replicationLag.trackUncommittedChange(changeDocument.clusterTime == null ? null : timestampToDate(changeDocument.clusterTime));
                             const transactionKeyValue = transactionKey(changeDocument);
                             if (transactionKeyValue == null || lastTxnKey != transactionKeyValue) {
@@ -856,31 +624,35 @@ export class ChangeStream {
                                 lastTxnKey = transactionKeyValue;
                                 transactionsReplicatedMetric.add(1);
                             }
-                            await this.writeChange(batch, table, changeDocument);
+                            for (const table of tablesToReplicate) {
+                                await this.writeChange(batch, table, changeDocument);
+                            }
                         }
                     }
                     else if (changeDocument.operationType == 'drop') {
-                        const rel = getMongoRelation(changeDocument.ns);
-                        const table = await this.getRelation(batch, rel, {
+                        const rel = getMongoRelation(changeDocument.ns, this.connections.connectionTag);
+                        const tables = await this.getRelations(batch, rel, {
                             // We're "dropping" this collection, so never snapshot it.
                             snapshot: false
                         });
-                        if (table.syncAny) {
-                            await batch.drop([table]);
-                            this.relationCache.delete(table);
+                        const tablesToDrop = tables.filter((table) => table.syncAny);
+                        if (tablesToDrop.length > 0) {
+                            await batch.drop(tablesToDrop);
                         }
+                        this.relationCache.delete(rel);
                     }
                     else if (changeDocument.operationType == 'rename') {
-                        const relFrom = getMongoRelation(changeDocument.ns);
-                        const relTo = getMongoRelation(changeDocument.to);
-                        const tableFrom = await this.getRelation(batch, relFrom, {
+                        const relFrom = getMongoRelation(changeDocument.ns, this.connections.connectionTag);
+                        const relTo = getMongoRelation(changeDocument.to, this.connections.connectionTag);
+                        const tablesFrom = await this.getRelations(batch, relFrom, {
                             // We're "dropping" this collection, so never snapshot it.
                             snapshot: false
                         });
-                        if (tableFrom.syncAny) {
-                            await batch.drop([tableFrom]);
-                            this.relationCache.delete(relFrom);
+                        const tablesToDrop = tablesFrom.filter((table) => table.syncAny);
+                        if (tablesToDrop.length > 0) {
+                            await batch.drop(tablesToDrop);
                         }
+                        this.relationCache.delete(relFrom);
                         // Here we do need to snapshot the new table
                         const collection = await this.getCollectionInfo(relTo.schema, relTo.name);
                         await this.handleRelation(batch, relTo, {
@@ -901,17 +673,18 @@ export class ChangeStream {
                     await batch.setResumeLsn(lsn);
                 }
                 batchSpan.end();
-                const durations = outerSpan.end();
-                const duration = batchSpan.endAt - batchSpan.startAt;
+                const durationsMicroseconds = outerSpan.end();
+                const duration = batchSpan.durationMillis;
                 this.logger.info(`Processed batch of ${events.length} changes / ${eventBatch.byteSize} bytes in ${duration}ms`, {
                     count: events.length,
                     bytes: eventBatch.byteSize,
                     duration,
-                    t: durations
+                    t: durationsMicroseconds
                 });
                 outerSpan = tracer.span('batch');
             }
         });
+        throw new ReplicationAbortedError(`Replication stream aborted`, this.abortSignal.reason);
     }
     getReplicationLagMillis() {
         return this.replicationLag.getLagMillis();
@@ -936,4 +709,26 @@ function transactionKey(doc) {
     }
     return `${doc.lsid.id.toString('hex')}:${doc.txnNumber}`;
 }
+/**
+ * Prioritize errors that are _not_ ReplicationAbortedError. Any error on either loopPromise or
+ * streamPromise aborts the other one, which then results in a ReplicationAbortedError, hiding the
+ * original cause.
+ */
+function replicationLoopError(results) {
+    // 1. Prioritize not ReplicationAbortedError.
+    for (const result of results) {
+        if (result.status == 'rejected' && !(result.reason instanceof ReplicationAbortedError)) {
+            return result.reason;
+        }
+    }
+    // 2. Fallback to ReplicationAbortedError.
+    for (const result of results) {
+        if (result.status == 'rejected') {
+            // At this point only ReplicationAbortedError remains
+            return result.reason;
+        }
+    }
+    // 3. Should never happen, but we cover this case.
+    return new ReplicationAssertionError(`Replication loop exited unexpectedly`);
+}
 //# sourceMappingURL=ChangeStream.js.map