@powersync/service-module-mongodb-storage 0.0.0-dev-20260225160713 → 0.0.0-dev-20260511080634

package/src/storage/implementation/MongoBucketBatch.ts

@@ -1,11 +1,10 @@
 import { mongo } from '@powersync/lib-service-mongodb';
-import { SqlEventDescriptor, SqliteRow, SqliteValue, HydratedSyncRules } from '@powersync/service-sync-rules';
+import { HydratedSyncRules, SqlEventDescriptor, SqliteRow, SqliteValue } from '@powersync/service-sync-rules';
 import * as bson from 'bson';
 
 import {
   BaseObserver,
   container,
-  logger as defaultLogger,
   ErrorCode,
   errors,
   Logger,
@@ -14,26 +13,27 @@ import {
 } from '@powersync/lib-services-framework';
 import {
   BucketStorageMarkRecordUnavailable,
+  CheckpointResult,
   deserializeBson,
   InternalOpId,
   isCompleteRow,
+  PerformanceTracer,
   SaveOperationTag,
   storage,
+  SyncRuleState,
   utils
 } from '@powersync/service-core';
 import * as timers from 'node:timers/promises';
-import { idPrefixFilter } from '../../utils/util.js';
-import { PowerSyncMongo } from './db.js';
-import { CurrentBucket, CurrentDataDocument, SourceKey, SyncRuleDocument } from './models.js';
+import { mongoTableId } from '../../utils/util.js';
+import { BucketDefinitionMapping } from './BucketDefinitionMapping.js';
+import { PersistedBatch } from './common/PersistedBatch.js';
+import { LoadedSourceRecord, SourceRecordStore } from './common/SourceRecordStore.js';
+import type { VersionedPowerSyncMongo } from './db.js';
+import { SyncRuleDocument } from './models.js';
+import { MAX_ROW_SIZE } from './MongoBucketBatchShared.js';
 import { MongoIdSequence } from './MongoIdSequence.js';
 import { batchCreateCustomWriteCheckpoints } from './MongoWriteCheckpointAPI.js';
-import { cacheKey, OperationBatch, RecordOperation } from './OperationBatch.js';
-import { PersistedBatch } from './PersistedBatch.js';
-
-/**
- * 15MB
- */
-export const MAX_ROW_SIZE = 15 * 1024 * 1024;
+import { OperationBatch, RecordOperation } from './OperationBatch.js';
 
 // Currently, we can only have a single flush() at a time, since it locks the op_id sequence.
 // While the MongoDB transaction retry mechanism handles this okay, using an in-process Mutex
@@ -43,15 +43,15 @@ export const MAX_ROW_SIZE = 15 * 1024 * 1024;
 const replicationMutex = new utils.Mutex();
 
 export interface MongoBucketBatchOptions {
-  db: PowerSyncMongo;
+  db: VersionedPowerSyncMongo;
   syncRules: HydratedSyncRules;
   groupId: number;
   slotName: string;
   lastCheckpointLsn: string | null;
   keepaliveOp: InternalOpId | null;
-  noCheckpointBeforeLsn: string;
   resumeFromLsn: string | null;
   storeCurrentData: boolean;
+  mapping: BucketDefinitionMapping;
   /**
    * Set to true for initial replication.
    */
@@ -59,31 +59,35 @@ export interface MongoBucketBatchOptions {
 
   markRecordUnavailable: BucketStorageMarkRecordUnavailable | undefined;
 
-  logger?: Logger;
+  logger: Logger;
+  tracer?: PerformanceTracer<'storage' | 'evaluate'>;
 }
 
-export class MongoBucketBatch
+export abstract class MongoBucketBatch
   extends BaseObserver<storage.BucketBatchStorageListener>
   implements storage.BucketStorageBatch
 {
-  private logger: Logger;
+  protected logger: Logger;
 
   private readonly client: mongo.MongoClient;
-  public readonly db: PowerSyncMongo;
+  public readonly db: VersionedPowerSyncMongo;
   public readonly session: mongo.ClientSession;
   private readonly sync_rules: HydratedSyncRules;
 
-  private readonly group_id: number;
+  protected readonly group_id: number;
 
   private readonly slot_name: string;
   private readonly storeCurrentData: boolean;
   private readonly skipExistingRows: boolean;
+  protected readonly mapping: BucketDefinitionMapping;
 
   private batch: OperationBatch | null = null;
   private write_checkpoint_batch: storage.CustomWriteCheckpointOptions[] = [];
   private markRecordUnavailable: BucketStorageMarkRecordUnavailable | undefined;
   private clearedError = false;
 
+  private tracer: PerformanceTracer<'storage' | 'evaluate'>;
+
   /**
    * Last LSN received associated with a checkpoint.
    *
@@ -93,8 +97,6 @@ export class MongoBucketBatch
    */
   private last_checkpoint_lsn: string | null = null;
 
-  private no_checkpoint_before_lsn: string;
-
   private persisted_op: InternalOpId | null = null;
 
   /**
@@ -118,22 +120,23 @@ export class MongoBucketBatch
 
   constructor(options: MongoBucketBatchOptions) {
     super();
-    this.logger = options.logger ?? defaultLogger;
+    this.logger = options.logger;
     this.client = options.db.client;
     this.db = options.db;
     this.group_id = options.groupId;
     this.last_checkpoint_lsn = options.lastCheckpointLsn;
-    this.no_checkpoint_before_lsn = options.noCheckpointBeforeLsn;
     this.resumeFromLsn = options.resumeFromLsn;
     this.session = this.client.startSession();
     this.slot_name = options.slotName;
     this.sync_rules = options.syncRules;
     this.storeCurrentData = options.storeCurrentData;
+    this.mapping = options.mapping;
     this.skipExistingRows = options.skipExistingRows;
     this.markRecordUnavailable = options.markRecordUnavailable;
     this.batch = new OperationBatch();
 
     this.persisted_op = options.keepaliveOp ?? null;
+    this.tracer = options.tracer ?? new PerformanceTracer('MongoDB storage');
   }
 
   addCustomWriteCheckpoint(checkpoint: storage.BatchedCustomWriteCheckpointOptions): void {
@@ -147,9 +150,11 @@ export class MongoBucketBatch
     return this.last_checkpoint_lsn;
   }
 
-  get noCheckpointBeforeLsn() {
-    return this.no_checkpoint_before_lsn;
-  }
+  protected abstract createPersistedBatch(writtenSize: number): PersistedBatch;
+
+  protected abstract get sourceRecordStore(): SourceRecordStore;
+
+  protected abstract cleanupDroppedSourceTables(sourceTables: storage.SourceTable[]): Promise<void>;
 
   async flush(options?: storage.BatchBucketFlushOptions): Promise<storage.FlushedResult | null> {
     let result: storage.FlushedResult | null = null;
@@ -169,6 +174,8 @@ export class MongoBucketBatch
     let last_op: InternalOpId | null = null;
     let resumeBatch: OperationBatch | null = null;
 
+    using _ = this.tracer.span('storage', 'flush');
+
     await this.withReplicationTransaction(`Flushing ${batch?.length ?? 0} ops`, async (session, opSeq) => {
       if (batch != null) {
         resumeBatch = await this.replicateBatch(session, batch, opSeq, options);
@@ -202,6 +209,7 @@ export class MongoBucketBatch
     options?: storage.BucketBatchCommitOptions
   ): Promise<OperationBatch | null> {
     let sizes: Map<string, number> | undefined = undefined;
+    using _ = this.tracer.span('storage', 'replicate_batch');
     if (this.storeCurrentData && !this.skipExistingRows) {
       // We skip this step if we don't store current_data, since the sizes will
       // always be small in that case.
@@ -216,32 +224,12 @@ export class MongoBucketBatch
       // (automatically limited to 48MB(?) per batch by MongoDB). The issue is that it changes
       // the order of processing, which then becomes really tricky to manage.
       // This now takes 2+ queries, but doesn't have any issues with order of operations.
-      const sizeLookups: SourceKey[] = batch.batch.map((r) => {
-        return { g: this.group_id, t: r.record.sourceTable.id, k: r.beforeId };
-      });
-
-      sizes = new Map<string, number>();
+      const sizeLookups = batch.batch.map((r) => ({
+        sourceTableId: mongoTableId(r.record.sourceTable.id),
+        replicaId: r.beforeId
+      }));
 
-      const sizeCursor: mongo.AggregationCursor<{ _id: SourceKey; size: number }> = this.db.current_data.aggregate(
-        [
-          {
-            $match: {
-              _id: { $in: sizeLookups }
-            }
-          },
-          {
-            $project: {
-              _id: 1,
-              size: { $bsonSize: '$$ROOT' }
-            }
-          }
-        ],
-        { session }
-      );
-      for await (let doc of sizeCursor.stream()) {
-        const key = cacheKey(doc._id.t, doc._id.k);
-        sizes.set(key, doc.size);
-      }
+      sizes = await this.sourceRecordStore.loadSizes(session, sizeLookups);
     }
 
     // If set, we need to start a new transaction with this batch.
@@ -259,64 +247,65 @@ export class MongoBucketBatch
         }
         continue;
       }
-      const lookups: SourceKey[] = b.map((r) => {
-        return { g: this.group_id, t: r.record.sourceTable.id, k: r.beforeId };
-      });
-      let current_data_lookup = new Map<string, CurrentDataDocument>();
-      // With skipExistingRows, we only need to know whether or not the row exists.
-      const projection = this.skipExistingRows ? { _id: 1 } : undefined;
-      const cursor = this.db.current_data.find(
-        {
-          _id: { $in: lookups }
-        },
-        { session, projection }
-      );
-      for await (let doc of cursor.stream()) {
-        current_data_lookup.set(cacheKey(doc._id.t, doc._id.k), doc);
-      }
-
-      let persistedBatch: PersistedBatch | null = new PersistedBatch(this.group_id, transactionSize, {
-        logger: this.logger
-      });
-
+      using lookupSpan = this.tracer.span('storage', 'lookup');
+      const lookups = b.map((r) => ({
+        sourceTableId: mongoTableId(r.record.sourceTable.id),
+        replicaId: r.beforeId
+      }));
+      let sourceRecordLookup = await this.sourceRecordStore.loadDocuments(session, lookups, this.skipExistingRows);
+      lookupSpan.end();
+
+      let persistedBatch: PersistedBatch | null = this.createPersistedBatch(transactionSize);
+
+      // The current code structure makes it tricky to cleanly split this span from the one
+      // where fluhsing. So we manually end and re-create this span whenever we flush.
+      let evalSpan = this.tracer.span('evaluate');
       for (let op of b) {
         if (resumeBatch) {
           resumeBatch.push(op);
           continue;
         }
-        const currentData = current_data_lookup.get(op.internalBeforeKey) ?? null;
-        if (currentData != null) {
+        const sourceRecord = sourceRecordLookup.get(op.internalBeforeKey) ?? null;
+        if (sourceRecord != null) {
           // If it will be used again later, it will be set again using nextData below
-          current_data_lookup.delete(op.internalBeforeKey);
+          sourceRecordLookup.delete(op.internalBeforeKey);
         }
-        const nextData = this.saveOperation(persistedBatch!, op, currentData, op_seq);
+        const nextData = this.saveOperation(persistedBatch!, op, sourceRecord, op_seq);
         if (nextData != null) {
           // Update our current_data and size cache
-          current_data_lookup.set(op.internalAfterKey!, nextData);
-          sizes?.set(op.internalAfterKey!, nextData.data.length());
+          sourceRecordLookup.set(op.internalAfterKey!, nextData);
+          sizes?.set(op.internalAfterKey!, nextData.data?.length() ?? 0);
         }
 
         if (persistedBatch!.shouldFlushTransaction()) {
+          evalSpan.end();
           // Transaction is getting big.
           // Flush, and resume in a new transaction.
-          const { flushedAny } = await persistedBatch!.flush(this.db, this.session, options);
+          using persistSpan = this.tracer.span('storage', 'persist_flush');
+          const { flushedAny } = await persistedBatch!.flush(this.session, options);
+
           didFlush ||= flushedAny;
           persistedBatch = null;
           // Computing our current progress is a little tricky here, since
           // we're stopping in the middle of a batch.
           // We create a new batch, and push any remaining operations to it.
           resumeBatch = new OperationBatch();
+          persistSpan.end();
+          evalSpan = this.tracer.span('evaluate');
         }
       }
+      evalSpan.end();
 
       if (persistedBatch) {
         transactionSize = persistedBatch.currentSize;
-        const { flushedAny } = await persistedBatch.flush(this.db, this.session, options);
+        using _ = this.tracer.span('storage', 'persist_flush');
+        const { flushedAny } = await persistedBatch.flush(this.session, options);
         didFlush ||= flushedAny;
       }
     }
 
     if (didFlush) {
+      using _ = this.tracer.span('storage', 'clear_error');
       await this.clearError();
     }
 
@@ -326,7 +315,7 @@ export class MongoBucketBatch
   private saveOperation(
     batch: PersistedBatch,
     operation: RecordOperation,
-    current_data: CurrentDataDocument | null,
+    sourceRecord: LoadedSourceRecord | null,
     opSeq: MongoIdSequence
   ) {
     const record = operation.record;
@@ -335,16 +324,16 @@ export class MongoBucketBatch
     let after = record.after;
     const sourceTable = record.sourceTable;
 
-    let existing_buckets: CurrentBucket[] = [];
-    let new_buckets: CurrentBucket[] = [];
-    let existing_lookups: bson.Binary[] = [];
-    let new_lookups: bson.Binary[] = [];
+    let existing_buckets: LoadedSourceRecord['buckets'] = [];
+    let new_buckets: LoadedSourceRecord['buckets'] = [];
+    let existing_lookups: LoadedSourceRecord['lookups'] = [];
+    let new_lookups: LoadedSourceRecord['lookups'] = [];
 
-    const before_key: SourceKey = { g: this.group_id, t: record.sourceTable.id, k: beforeId };
+    const sourceTableId = mongoTableId(record.sourceTable.id);
 
     if (this.skipExistingRows) {
       if (record.tag == SaveOperationTag.INSERT) {
-        if (current_data != null) {
+        if (sourceRecord != null) {
           // Initial replication, and we already have the record.
           // This may be a different version of the record, but streaming replication
           // will take care of that.
@@ -357,7 +346,7 @@ export class MongoBucketBatch
     }
 
     if (record.tag == SaveOperationTag.UPDATE) {
-      const result = current_data;
+      const result = sourceRecord;
       if (result == null) {
         // Not an error if we re-apply a transaction
         existing_buckets = [];
@@ -378,13 +367,13 @@ export class MongoBucketBatch
       } else {
         existing_buckets = result.buckets;
         existing_lookups = result.lookups;
-        if (this.storeCurrentData) {
-          const data = deserializeBson((result.data as mongo.Binary).buffer) as SqliteRow;
+        if (this.storeCurrentData && result.data != null) {
+          const data = deserializeBson(result.data.buffer) as SqliteRow;
           after = storage.mergeToast<SqliteValue>(after!, data);
         }
       }
     } else if (record.tag == SaveOperationTag.DELETE) {
-      const result = current_data;
+      const result = sourceRecord;
       if (result == null) {
         // Not an error if we re-apply a transaction
         existing_buckets = [];
@@ -401,9 +390,9 @@ export class MongoBucketBatch
       }
     }
 
-    let afterData: bson.Binary | undefined;
+    let afterData: bson.Binary | null = null;
     if (afterId != null && !this.storeCurrentData) {
-      afterData = new bson.Binary(bson.serialize({}));
+      afterData = null;
     } else if (afterId != null) {
       try {
         // This will fail immediately if the record is > 16MB.
@@ -501,13 +490,7 @@ export class MongoBucketBatch
           table: sourceTable,
           before_buckets: existing_buckets
         });
-        new_buckets = evaluated.map((e) => {
-          return {
-            bucket: e.bucket,
-            table: e.table,
-            id: e.id
-          };
-        });
+        new_buckets = this.sourceRecordStore.mapEvaluatedBuckets(evaluated);
       }
 
       if (sourceTable.syncParameters) {
@@ -540,40 +523,46 @@ export class MongoBucketBatch
           evaluated: paramEvaluated,
           existing_lookups
         });
-        new_lookups = paramEvaluated.map((p) => {
-          return storage.serializeLookup(p.lookup);
-        });
+        new_lookups = this.sourceRecordStore.mapParameterLookups(paramEvaluated);
       }
     }
 
-    let result: CurrentDataDocument | null = null;
+    let result: LoadedSourceRecord | null = null;
 
     // 5. TOAST: Update current data and bucket list.
     if (afterId) {
       // Insert or update
-      const after_key: SourceKey = { g: this.group_id, t: sourceTable.id, k: afterId };
-      batch.upsertCurrentData(after_key, {
+      batch.upsertCurrentData({
+        sourceTableId,
+        replicaId: afterId,
         data: afterData,
         buckets: new_buckets,
         lookups: new_lookups
       });
       result = {
-        _id: after_key,
-        data: afterData!,
+        sourceTableId,
+        replicaId: afterId,
+        data: afterData,
         buckets: new_buckets,
-        lookups: new_lookups
+        lookups: new_lookups,
+        cacheKey: operation.internalAfterKey!
       };
     }
 
     if (afterId == null || !storage.replicaIdEquals(beforeId, afterId)) {
       // Either a delete (afterId == null), or replaced the old replication id
-      batch.deleteCurrentData(before_key);
+      // Note that this is a soft delete.
+      // We don't specifically need a new or unique op_id here, but it must be greater than the
+      // last checkpoint, so we use next().
+      batch.softDeleteCurrentData(sourceTableId, beforeId, opSeq.next());
     }
     return result;
   }
 
   private async withTransaction(cb: () => Promise<void>) {
+    using lockSpan = this.tracer.span('storage', 'internal_lock');
     await replicationMutex.exclusiveLock(async () => {
+      lockSpan.end();
       await this.session.withTransaction(
         async () => {
           try {
@@ -584,7 +573,9 @@ export class MongoBucketBatch
             } else {
               this.logger.warn('Transaction error', e as Error);
             }
-            await timers.setTimeout(Math.random() * 50);
+            const delay = Math.random() * 50;
+            using _ = this.tracer.span('storage', 'retry_delay');
+            await timers.setTimeout(delay);
             throw e;
           }
         },
@@ -664,75 +655,28 @@ export class MongoBucketBatch
   }
 
   async [Symbol.asyncDispose]() {
+    if (this.batch != null || this.write_checkpoint_batch.length > 0) {
+      // We don't error here, since:
+      // 1. In error states, this is expected (we can't distinguish between disposing after success or error).
+      // 2. SuppressedError is messy to deal with.
+      this.logger.warn('Disposing writer with unflushed changes');
+    }
     await this.session.endSession();
     super.clearListeners();
   }
 
+  async dispose() {
+    await this[Symbol.asyncDispose]();
+  }
+
   private lastWaitingLogThottled = 0;
 
-  async commit(lsn: string, options?: storage.BucketBatchCommitOptions): Promise<boolean> {
+  async commit(lsn: string, options?: storage.BucketBatchCommitOptions): Promise<CheckpointResult> {
     const { createEmptyCheckpoints } = { ...storage.DEFAULT_BUCKET_BATCH_COMMIT_OPTIONS, ...options };
 
     await this.flush(options);
 
-    if (this.last_checkpoint_lsn != null && lsn < this.last_checkpoint_lsn) {
-      // When re-applying transactions, don't create a new checkpoint until
-      // we are past the last transaction.
-      this.logger.info(`Re-applied transaction ${lsn} - skipping checkpoint`);
-      // Cannot create a checkpoint yet - return false
-      return false;
-    }
-    if (lsn < this.no_checkpoint_before_lsn) {
-      if (Date.now() - this.lastWaitingLogThottled > 5_000) {
-        this.logger.info(
-          `Waiting until ${this.no_checkpoint_before_lsn} before creating checkpoint, currently at ${lsn}. Persisted op: ${this.persisted_op}`
-        );
-        this.lastWaitingLogThottled = Date.now();
-      }
-
-      // Edge case: During initial replication, we have a no_checkpoint_before_lsn set,
-      // and don't actually commit the snapshot.
-      // The first commit can happen from an implicit keepalive message.
-      // That needs the persisted_op to get an accurate checkpoint, so
-      // we persist that in keepalive_op.
-
-      await this.db.sync_rules.updateOne(
-        {
-          _id: this.group_id
-        },
-        {
-          $set: {
-            keepalive_op: this.persisted_op == null ? null : String(this.persisted_op)
-          }
-        },
-        { session: this.session }
-      );
-      await this.db.notifyCheckpoint();
-
-      // Cannot create a checkpoint yet - return false
-      return false;
-    }
-
-    if (!createEmptyCheckpoints && this.persisted_op == null) {
-      // Nothing to commit - also return true
-      await this.autoActivate(lsn);
-      return true;
-    }
-
     const now = new Date();
-    const update: Partial<SyncRuleDocument> = {
-      last_checkpoint_lsn: lsn,
-      last_checkpoint_ts: now,
-      last_keepalive_ts: now,
-      snapshot_done: true,
-      last_fatal_error: null,
-      last_fatal_error_ts: null,
-      keepalive_op: null
-    };
-
-    if (this.persisted_op != null) {
-      update.last_checkpoint = this.persisted_op;
-    }
 
     // Mark relevant write checkpoints as "processed".
     // This makes it easier to identify write checkpoints that are "valid" in order.
@@ -751,21 +695,155 @@ export class MongoBucketBatch
       }
     );
 
-    await this.db.sync_rules.updateOne(
-      {
-        _id: this.group_id
-      },
+    const can_checkpoint = {
+      $and: [
+        { $eq: ['$snapshot_done', true] },
+        {
+          $or: [{ $eq: ['$last_checkpoint_lsn', null] }, { $lte: ['$last_checkpoint_lsn', { $literal: lsn }] }]
+        },
+        {
+          $or: [{ $eq: ['$no_checkpoint_before', null] }, { $lte: ['$no_checkpoint_before', { $literal: lsn }] }]
+        }
+      ]
+    };
+
+    const new_keepalive_op = {
+      $cond: [
+        can_checkpoint,
+        { $literal: null },
+        {
+          $toString: {
+            $max: [{ $toLong: '$keepalive_op' }, { $literal: this.persisted_op }, 0n]
+          }
+        }
+      ]
+    };
+
+    const new_last_checkpoint = {
+      $cond: [
+        can_checkpoint,
+        {
+          $max: ['$last_checkpoint', { $literal: this.persisted_op }, { $toLong: '$keepalive_op' }, 0n]
+        },
+        '$last_checkpoint'
+      ]
+    };
+
+    // For this query, we need to handle multiple cases, depending on the state:
+    // 1. Normal commit - advance last_checkpoint to this.persisted_op.
+    // 2. Commit delayed by no_checkpoint_before due to snapshot. In this case we only advance keepalive_op.
+    // 3. Commit with no new data - here may may set last_checkpoint = keepalive_op, if a delayed commit is relevant.
+    // We want to do as much as possible in a single atomic database operation, which makes this somewhat complex.
+    let preUpdateDocument = await this.db.sync_rules.findOneAndUpdate(
+      { _id: this.group_id },
+      [
+        {
+          $set: {
+            _can_checkpoint: can_checkpoint,
+            _not_empty: createEmptyCheckpoints
+              ? true
+              : {
+                  $or: [
+                    { $literal: createEmptyCheckpoints },
+                    { $ne: ['$keepalive_op', new_keepalive_op] },
+                    { $ne: ['$last_checkpoint', new_last_checkpoint] }
+                  ]
+                }
+          }
+        },
+        {
+          $set: {
+            last_checkpoint_lsn: {
+              $cond: [{ $and: ['$_can_checkpoint', '$_not_empty'] }, { $literal: lsn }, '$last_checkpoint_lsn']
+            },
+            last_checkpoint_ts: {
+              $cond: [{ $and: ['$_can_checkpoint', '$_not_empty'] }, { $literal: now }, '$last_checkpoint_ts']
+            },
+            last_keepalive_ts: { $literal: now },
+            last_fatal_error: { $literal: null },
+            last_fatal_error_ts: { $literal: null },
+            keepalive_op: new_keepalive_op,
+            last_checkpoint: new_last_checkpoint,
+            // Unset snapshot_lsn on checkpoint
+            snapshot_lsn: {
+              $cond: [{ $and: ['$_can_checkpoint', '$_not_empty'] }, { $literal: null }, '$snapshot_lsn']
+            }
+          }
+        },
+        {
+          $unset: ['_can_checkpoint', '_not_empty']
+        }
+      ],
       {
-        $set: update,
-        $unset: { snapshot_lsn: 1 }
-      },
-      { session: this.session }
+        session: this.session,
+        // We return the before document, so that we can check the previous state to determine if a checkpoint was actually created or if we were blocked by snapshot/no_checkpoint_before.
+        returnDocument: 'before',
+        projection: {
+          snapshot_done: 1,
+          last_checkpoint_lsn: 1,
+          no_checkpoint_before: 1,
+          keepalive_op: 1,
+          last_checkpoint: 1
+        }
+      }
     );
-    await this.autoActivate(lsn);
-    await this.db.notifyCheckpoint();
-    this.persisted_op = null;
-    this.last_checkpoint_lsn = lsn;
-    return true;
+
+    if (preUpdateDocument == null) {
+      throw new ReplicationAssertionError(
+        'Failed to update checkpoint - no matching sync_rules document for _id: ' + this.group_id
+      );
+    }
+
+    // This re-implements the same logic as in the pipeline, to determine what was actually updated.
+    // Unfortunately we cannot return these from the pipeline directly, so we need to re-implement the logic.
+    const canCheckpoint =
+      preUpdateDocument.snapshot_done === true &&
+      (preUpdateDocument.last_checkpoint_lsn == null || preUpdateDocument.last_checkpoint_lsn <= lsn) &&
+      (preUpdateDocument.no_checkpoint_before == null || preUpdateDocument.no_checkpoint_before <= lsn);
+
+    const keepaliveOp = preUpdateDocument.keepalive_op == null ? null : BigInt(preUpdateDocument.keepalive_op);
+    const maxKeepalive = [keepaliveOp ?? 0n, this.persisted_op ?? 0n, 0n].reduce((a, b) => (a > b ? a : b));
+    const newKeepaliveOp = canCheckpoint ? null : maxKeepalive.toString();
+    const newLastCheckpoint = canCheckpoint
+      ? [preUpdateDocument.last_checkpoint ?? 0n, this.persisted_op ?? 0n, keepaliveOp ?? 0n, 0n].reduce((a, b) =>
+          a > b ? a : b
+        )
+      : preUpdateDocument.last_checkpoint;
+    const notEmpty =
+      createEmptyCheckpoints ||
+      preUpdateDocument.keepalive_op !== newKeepaliveOp ||
+      preUpdateDocument.last_checkpoint !== newLastCheckpoint;
+    const checkpointCreated = canCheckpoint && notEmpty;
+
+    const checkpointBlocked = !canCheckpoint;
+
+    if (checkpointBlocked) {
+      // Failed on snapshot_done or no_checkpoint_before.
+      if (Date.now() - this.lastWaitingLogThottled > 5_000) {
+        this.logger.info(
+          `Waiting before creating checkpoint, currently at ${lsn} / ${preUpdateDocument.keepalive_op}. Current state: ${JSON.stringify(
+            {
+              snapshot_done: preUpdateDocument.snapshot_done,
+              last_checkpoint_lsn: preUpdateDocument.last_checkpoint_lsn,
+              no_checkpoint_before: preUpdateDocument.no_checkpoint_before
+            }
+          )}`
+        );
+        this.lastWaitingLogThottled = Date.now();
+      }
+    } else {
+      if (checkpointCreated) {
+        this.logger.debug(`Created checkpoint at ${lsn} / ${newLastCheckpoint}`);
+      }
+      await this.autoActivate(lsn);
+      await this.db.notifyCheckpoint();
+      this.persisted_op = null;
+      this.last_checkpoint_lsn = lsn;
+      if (newLastCheckpoint != null) {
+        await this.sourceRecordStore.postCommitCleanup(newLastCheckpoint, this.logger);
+      }
+    }
+    return { checkpointBlocked, checkpointCreated };
   }
 
   /**
@@ -785,7 +863,7 @@ export class MongoBucketBatch
     let activated = false;
     await session.withTransaction(async () => {
       const doc = await this.db.sync_rules.findOne({ _id: this.group_id }, { session });
-      if (doc && doc.state == 'PROCESSING') {
+      if (doc && doc.state == SyncRuleState.PROCESSING && doc.snapshot_done && doc.last_checkpoint != null) {
         await this.db.sync_rules.updateOne(
           {
             _id: this.group_id
@@ -811,68 +889,19 @@ export class MongoBucketBatch
           { session }
         );
         activated = true;
+      } else if (doc?.state != SyncRuleState.PROCESSING) {
+        this.needsActivation = false;
       }
     });
     if (activated) {
-      this.logger.info(`Activated new sync rules at ${lsn}`);
+      this.logger.info(`Activated new replication stream at ${lsn}`);
       await this.db.notifyCheckpoint();
+      this.needsActivation = false;
     }
-    this.needsActivation = false;
   }
 
-  async keepalive(lsn: string): Promise<boolean> {
-    if (this.last_checkpoint_lsn != null && lsn < this.last_checkpoint_lsn) {
-      // No-op
-      return false;
-    }
-
-    if (lsn < this.no_checkpoint_before_lsn) {
-      return false;
-    }
-
-    if (this.persisted_op != null) {
-      // The commit may have been skipped due to "no_checkpoint_before_lsn".
-      // Apply it now if relevant
-      this.logger.info(`Commit due to keepalive at ${lsn} / ${this.persisted_op}`);
-      return await this.commit(lsn);
-    }
-
-    await this.db.write_checkpoints.updateMany(
-      {
-        processed_at_lsn: null,
-        'lsns.1': { $lte: lsn }
-      },
-      {
-        $set: {
-          processed_at_lsn: lsn
-        }
-      },
-      {
-        session: this.session
-      }
-    );
-
-    await this.db.sync_rules.updateOne(
-      {
-        _id: this.group_id
-      },
-      {
-        $set: {
-          last_checkpoint_lsn: lsn,
-          snapshot_done: true,
-          last_fatal_error: null,
-          last_fatal_error_ts: null,
-          last_keepalive_ts: new Date()
-        },
-        $unset: { snapshot_lsn: 1 }
-      },
-      { session: this.session }
-    );
-    await this.autoActivate(lsn);
-    await this.db.notifyCheckpoint();
-    this.last_checkpoint_lsn = lsn;
-
-    return true;
+  async keepalive(lsn: string): Promise<CheckpointResult> {
+    return await this.commit(lsn, { createEmptyCheckpoints: true });
   }
 
   async setResumeLsn(lsn: string): Promise<void> {
@@ -938,9 +967,11 @@ export class MongoBucketBatch
 
     await this.withTransaction(async () => {
       for (let table of sourceTables) {
-        await this.db.source_tables.deleteOne({ _id: table.id });
+        await this.db.commonSourceTables(this.group_id).deleteOne({ _id: mongoTableId(table.id) });
       }
     });
+
+    await this.cleanupDroppedSourceTables(sourceTables);
     return result;
   }
 
@@ -972,21 +1003,10 @@ export class MongoBucketBatch
     let lastBatchCount = BATCH_LIMIT;
     while (lastBatchCount == BATCH_LIMIT) {
       await this.withReplicationTransaction(`Truncate ${sourceTable.qualifiedName}`, async (session, opSeq) => {
-        const current_data_filter: mongo.Filter<CurrentDataDocument> = {
-          _id: idPrefixFilter<SourceKey>({ g: this.group_id, t: sourceTable.id }, ['k'])
-        };
-
-        const cursor = this.db.current_data.find(current_data_filter, {
-          projection: {
-            _id: 1,
-            buckets: 1,
-            lookups: 1
-          },
-          limit: BATCH_LIMIT,
-          session: session
-        });
-        const batch = await cursor.toArray();
-        const persistedBatch = new PersistedBatch(this.group_id, 0, { logger: this.logger });
+        using evalSpan = this.tracer.span('evaluate');
+        const sourceTableId = mongoTableId(sourceTable.id);
+        const batch = await this.sourceRecordStore.loadTruncateBatch(session, sourceTableId, BATCH_LIMIT);
+        const persistedBatch = this.createPersistedBatch(0);
 
         for (let value of batch) {
           persistedBatch.saveBucketData({
@@ -994,19 +1014,23 @@ export class MongoBucketBatch
             before_buckets: value.buckets,
             evaluated: [],
             table: sourceTable,
-            sourceKey: value._id.k
+            sourceKey: value.replicaId
           });
           persistedBatch.saveParameterData({
             op_seq: opSeq,
             existing_lookups: value.lookups,
             evaluated: [],
             sourceTable: sourceTable,
-            sourceKey: value._id.k
+            sourceKey: value.replicaId
           });
 
-          persistedBatch.deleteCurrentData(value._id);
+          // Since this is not from streaming replication, we can do a hard delete
+          persistedBatch.hardDeleteCurrentData(sourceTableId, value.replicaId);
         }
-        await persistedBatch.flush(this.db, session);
+        evalSpan.end();
+
+        using _ = this.tracer.span('storage', 'persist_flush');
+        await persistedBatch.flush(session);
         lastBatchCount = batch.length;
 
         last_op = opSeq.last();
@@ -1029,8 +1053,8 @@ export class MongoBucketBatch
     copy.snapshotStatus = snapshotStatus;
 
     await this.withTransaction(async () => {
-      await this.db.source_tables.updateOne(
-        { _id: table.id },
+      await this.db.commonSourceTables(this.group_id).updateOne(
+        { _id: mongoTableId(table.id) },
         {
           $set: {
             snapshot_status: {
@@ -1047,12 +1071,44 @@ export class MongoBucketBatch
     return copy;
   }
 
-  async markSnapshotDone(tables: storage.SourceTable[], no_checkpoint_before_lsn: string) {
+  async markAllSnapshotDone(no_checkpoint_before_lsn: string) {
+    await this.db.sync_rules.updateOne(
+      {
+        _id: this.group_id
+      },
+      {
+        $set: {
+          snapshot_done: true,
+          last_keepalive_ts: new Date()
+        },
+        $max: {
+          no_checkpoint_before: no_checkpoint_before_lsn
+        }
+      },
+      { session: this.session }
+    );
+  }
+
+  async markTableSnapshotRequired(table: storage.SourceTable): Promise<void> {
+    await this.db.sync_rules.updateOne(
+      {
+        _id: this.group_id
+      },
+      {
+        $set: {
+          snapshot_done: false
+        }
+      },
+      { session: this.session }
+    );
+  }
+
+  async markTableSnapshotDone(tables: storage.SourceTable[], no_checkpoint_before_lsn?: string) {
     const session = this.session;
-    const ids = tables.map((table) => table.id);
+    const ids = tables.map((table) => mongoTableId(table.id));
 
     await this.withTransaction(async () => {
-      await this.db.source_tables.updateMany(
+      await this.db.commonSourceTables(this.group_id).updateMany(
         { _id: { $in: ids } },
         {
           $set: {
@@ -1065,17 +1121,17 @@ export class MongoBucketBatch
         { session }
       );
 
-      if (no_checkpoint_before_lsn > this.no_checkpoint_before_lsn) {
-        this.no_checkpoint_before_lsn = no_checkpoint_before_lsn;
-
+      if (no_checkpoint_before_lsn != null) {
         await this.db.sync_rules.updateOne(
           {
             _id: this.group_id
           },
           {
             $set: {
-              no_checkpoint_before: no_checkpoint_before_lsn,
               last_keepalive_ts: new Date()
+            },
+            $max: {
+              no_checkpoint_before: no_checkpoint_before_lsn
             }
           },
           { session: this.session }
@@ -1118,7 +1174,3 @@ export class MongoBucketBatch
     );
   }
 }
-
-export function currentBucketKey(b: CurrentBucket) {
-  return `${b.bucket}/${b.table}/${b.id}`;
-}