@powersync/service-core 0.0.0-dev-20240718134716 → 0.0.0-dev-20240725112650

package/src/storage/mongo/MongoCompactor.ts

@@ -0,0 +1,371 @@
+import { logger } from '@powersync/lib-services-framework';
+import { AnyBulkWriteOperation, MaxKey, MinKey } from 'mongodb';
+import { addChecksums } from '../../util/utils.js';
+import { PowerSyncMongo } from './db.js';
+import { BucketDataDocument, BucketDataKey } from './models.js';
+import { CompactOptions } from '../BucketStorage.js';
+
+interface CurrentBucketState {
+  /** Bucket name */
+  bucket: string;
+  /**
+   * Rows seen in the bucket, with the last op_id of each.
+   */
+  seen: Map<string, bigint>;
+  /**
+   * Estimated memory usage of the seen Map.
+   */
+  trackingSize: number;
+
+  /**
+   * Last (lowest) seen op_id that is not a PUT.
+   */
+  lastNotPut: bigint | null;
+
+  /**
+   * Number of REMOVE/MOVE operations seen since lastNotPut.
+   */
+  opsSincePut: number;
+}
+
+/**
+ * Additional options, primarily for testing.
+ */
+export interface MongoCompactOptions extends CompactOptions {
+  /** Minimum of 2 */
+  clearBatchLimit?: number;
+  /** Minimum of 1 */
+  moveBatchLimit?: number;
+  /** Minimum of 1 */
+  moveBatchQueryLimit?: number;
+}
+
+const DEFAULT_CLEAR_BATCH_LIMIT = 5000;
+const DEFAULT_MOVE_BATCH_LIMIT = 2000;
+const DEFAULT_MOVE_BATCH_QUERY_LIMIT = 10_000;
+
+/** This default is primarily for tests. */
+const DEFAULT_MEMORY_LIMIT_MB = 64;
+
+export class MongoCompactor {
+  private updates: AnyBulkWriteOperation<BucketDataDocument>[] = [];
+
+  private idLimitBytes: number;
+  private moveBatchLimit: number;
+  private moveBatchQueryLimit: number;
+  private clearBatchLimit: number;
+  private maxOpId: bigint | undefined;
+  private buckets: string[] | undefined;
+
+  constructor(private db: PowerSyncMongo, private group_id: number, options?: MongoCompactOptions) {
+    this.idLimitBytes = (options?.memoryLimitMB ?? DEFAULT_MEMORY_LIMIT_MB) * 1024 * 1024;
+    this.moveBatchLimit = options?.moveBatchLimit ?? DEFAULT_MOVE_BATCH_LIMIT;
+    this.moveBatchQueryLimit = options?.moveBatchQueryLimit ?? DEFAULT_MOVE_BATCH_QUERY_LIMIT;
+    this.clearBatchLimit = options?.clearBatchLimit ?? DEFAULT_CLEAR_BATCH_LIMIT;
+    this.maxOpId = options?.maxOpId;
+    this.buckets = options?.compactBuckets;
+  }
+
+  /**
+   * Compact buckets by converting operations into MOVE and/or CLEAR operations.
+   *
+   * See /docs/compacting-operations.md for details.
+   */
+  async compact() {
+    if (this.buckets) {
+      for (let bucket of this.buckets) {
+        // We can make this more efficient later on by iterating
+        // through the buckets in a single query.
+        // That makes batching more tricky, so we leave for later.
+        await this.compactInternal(bucket);
+      }
+    } else {
+      await this.compactInternal(undefined);
+    }
+  }
+
+  async compactInternal(bucket: string | undefined) {
+    const idLimitBytes = this.idLimitBytes;
+
+    let currentState: CurrentBucketState | null = null;
+
+    // Constant lower bound
+    const lowerBound: BucketDataKey = {
+      g: this.group_id,
+      b: bucket ?? (new MinKey() as any),
+      o: new MinKey() as any
+    };
+
+    // Upper bound is adjusted for each batch
+    let upperBound: BucketDataKey = {
+      g: this.group_id,
+      b: bucket ?? (new MaxKey() as any),
+      o: new MaxKey() as any
+    };
+
+    while (true) {
+      // Query one batch at a time, to avoid cursor timeouts
+      const batch = await this.db.bucket_data
+        .find(
+          {
+            _id: {
+              $gte: lowerBound,
+              $lt: upperBound
+            }
+          },
+          {
+            projection: {
+              _id: 1,
+              op: 1,
+              table: 1,
+              row_id: 1,
+              source_table: 1,
+              source_key: 1
+            },
+            limit: this.moveBatchQueryLimit,
+            sort: { _id: -1 },
+            singleBatch: true
+          }
+        )
+        .toArray();
+
+      if (batch.length == 0) {
+        // We've reached the end
+        break;
+      }
+
+      // Set upperBound for the next batch
+      upperBound = batch[batch.length - 1]._id;
+
+      for (let doc of batch) {
+        if (currentState == null || doc._id.b != currentState.bucket) {
+          if (currentState != null && currentState.lastNotPut != null && currentState.opsSincePut >= 1) {
+            // Important to flush before clearBucket()
+            await this.flush();
+            logger.info(
+              `Inserting CLEAR at ${this.group_id}:${currentState.bucket}:${currentState.lastNotPut} to remove ${currentState.opsSincePut} operations`
+            );
+
+            const bucket = currentState.bucket;
+            const clearOp = currentState.lastNotPut;
+            // Free memory before clearing bucket
+            currentState = null;
+            await this.clearBucket(bucket, clearOp);
+          }
+          currentState = {
+            bucket: doc._id.b,
+            seen: new Map(),
+            trackingSize: 0,
+            lastNotPut: null,
+            opsSincePut: 0
+          };
+        }
+
+        if (this.maxOpId != null && doc._id.o > this.maxOpId) {
+          continue;
+        }
+
+        let isPersistentPut = doc.op == 'PUT';
+
+        if (doc.op == 'REMOVE' || doc.op == 'PUT') {
+          const key = `${doc.table}/${doc.row_id}/${doc.source_table}/${doc.source_key?.toHexString()}`;
+          const targetOp = currentState.seen.get(key);
+          if (targetOp) {
+            // Will convert to MOVE, so don't count as PUT
+            isPersistentPut = false;
+
+            this.updates.push({
+              updateOne: {
+                filter: {
+                  _id: doc._id
+                },
+                update: {
+                  $set: {
+                    op: 'MOVE',
+                    target_op: targetOp
+                  },
+                  $unset: {
+                    source_table: 1,
+                    source_key: 1,
+                    table: 1,
+                    row_id: 1,
+                    data: 1
+                  }
+                }
+              }
+            });
+          } else {
+            if (currentState.trackingSize >= idLimitBytes) {
+              // Reached memory limit.
+              // Keep the highest seen values in this case.
+            } else {
+              // flatstr reduces the memory usage by flattening the string
+              currentState.seen.set(flatstr(key), doc._id.o);
+              // length + 16 for the string
+              // 24 for the bigint
+              // 50 for map overhead
+              // 50 for additional overhead
+              currentState.trackingSize += key.length + 140;
+            }
+          }
+        }
+
+        if (isPersistentPut) {
+          currentState.lastNotPut = null;
+          currentState.opsSincePut = 0;
+        } else if (doc.op != 'CLEAR') {
+          if (currentState.lastNotPut == null) {
+            currentState.lastNotPut = doc._id.o;
+          }
+          currentState.opsSincePut += 1;
+        }
+
+        if (this.updates.length >= this.moveBatchLimit) {
+          await this.flush();
+        }
+      }
+    }
+
+    await this.flush();
+    currentState?.seen.clear();
+    if (currentState?.lastNotPut != null && currentState?.opsSincePut > 1) {
+      logger.info(
+        `Inserting CLEAR at ${this.group_id}:${currentState.bucket}:${currentState.lastNotPut} to remove ${currentState.opsSincePut} operations`
+      );
+      const bucket = currentState.bucket;
+      const clearOp = currentState.lastNotPut;
+      // Free memory before clearing bucket
+      currentState = null;
+      await this.clearBucket(bucket, clearOp);
+    }
+  }
+
+  private async flush() {
+    if (this.updates.length > 0) {
+      logger.info(`Compacting ${this.updates.length} ops`);
+      await this.db.bucket_data.bulkWrite(this.updates, {
+        // Order is not important.
+        // Since checksums are not affected, these operations can happen in any order,
+        // and it's fine if the operations are partially applied.
+        // Each individual operation is atomic.
+        ordered: false
+      });
+      this.updates = [];
+    }
+  }
+
+  /**
+   * Perform a CLEAR compact for a bucket.
+   *
+   * @param bucket bucket name
+   * @param op op_id of the last non-PUT operation, which will be converted to CLEAR.
+   */
+  private async clearBucket(bucket: string, op: bigint) {
+    const opFilter = {
+      _id: {
+        $gte: {
+          g: this.group_id,
+          b: bucket,
+          o: new MinKey() as any
+        },
+        $lte: {
+          g: this.group_id,
+          b: bucket,
+          o: op
+        }
+      }
+    };
+
+    const session = this.db.client.startSession();
+    try {
+      let done = false;
+      while (!done) {
+        // Do the CLEAR operation in batches, with each batch a separate transaction.
+        // The state after each batch is fully consistent.
+        // We need a transaction per batch to make sure checksums stay consistent.
+        await session.withTransaction(
+          async () => {
+            const query = this.db.bucket_data.find(opFilter, {
+              session,
+              sort: { _id: 1 },
+              projection: {
+                _id: 1,
+                op: 1,
+                checksum: 1,
+                target_op: 1
+              },
+              limit: this.clearBatchLimit
+            });
+            let checksum = 0;
+            let lastOpId: BucketDataKey | null = null;
+            let targetOp: bigint | null = null;
+            let gotAnOp = false;
+            for await (let op of query.stream()) {
+              if (op.op == 'MOVE' || op.op == 'REMOVE' || op.op == 'CLEAR') {
+                checksum = addChecksums(checksum, op.checksum);
+                lastOpId = op._id;
+                if (op.op != 'CLEAR') {
+                  gotAnOp = true;
+                }
+                if (op.target_op != null) {
+                  if (targetOp == null || op.target_op > targetOp) {
+                    targetOp = op.target_op;
+                  }
+                }
+              } else {
+                throw new Error(`Unexpected ${op.op} operation at ${op._id.g}:${op._id.b}:${op._id.o}`);
+              }
+            }
+            if (!gotAnOp) {
+              done = true;
+              return;
+            }
+
+            logger.info(`Flushing CLEAR at ${lastOpId?.o}`);
+            await this.db.bucket_data.deleteMany(
+              {
+                _id: {
+                  $gte: {
+                    g: this.group_id,
+                    b: bucket,
+                    o: new MinKey() as any
+                  },
+                  $lte: lastOpId!
+                }
+              },
+              { session }
+            );
+
+            await this.db.bucket_data.insertOne(
+              {
+                _id: lastOpId!,
+                op: 'CLEAR',
+                checksum: checksum,
+                data: null,
+                target_op: targetOp
+              },
+              { session }
+            );
+          },
+          {
+            writeConcern: { w: 'majority' },
+            readConcern: { level: 'snapshot' }
+          }
+        );
+      }
+    } finally {
+      await session.endSession();
+    }
+  }
+}
+
+/**
+ * Flattens string to reduce memory usage (around 320 bytes -> 120 bytes),
+ * at the cost of some upfront CPU usage.
+ *
+ * From: https://github.com/davidmarkclements/flatstr/issues/8
+ */
+function flatstr(s: string) {
+  s.match(/\n/g);
+  return s;
+}

package/src/storage/mongo/MongoSyncBucketStorage.ts

@@ -8,21 +8,24 @@ import * as util from '../../util/util-index.js';
 import {
   BucketDataBatchOptions,
   BucketStorageBatch,
+  CompactOptions,
   DEFAULT_DOCUMENT_BATCH_LIMIT,
   DEFAULT_DOCUMENT_CHUNK_LIMIT_BYTES,
   FlushedResult,
   ResolveTableOptions,
   ResolveTableResult,
+  SyncBucketDataBatch,
   SyncRulesBucketStorage,
   SyncRuleStatus
 } from '../BucketStorage.js';
+import { ChecksumCache, FetchPartialBucketChecksum } from '../ChecksumCache.js';
 import { MongoBucketStorage } from '../MongoBucketStorage.js';
 import { SourceTable } from '../SourceTable.js';
 import { PowerSyncMongo } from './db.js';
 import { BucketDataDocument, BucketDataKey, SourceKey, SyncRuleState } from './models.js';
 import { MongoBucketBatch } from './MongoBucketBatch.js';
-import { BSON_DESERIALIZE_OPTIONS, idPrefixFilter, readSingleBatch, serializeLookup } from './util.js';
-import { ChecksumCache, FetchPartialBucketChecksum } from '../ChecksumCache.js';
+import { MongoCompactor } from './MongoCompactor.js';
+import { BSON_DESERIALIZE_OPTIONS, idPrefixFilter, mapOpEntry, readSingleBatch, serializeLookup } from './util.js';
 
 export class MongoSyncBucketStorage implements SyncRulesBucketStorage {
   private readonly db: PowerSyncMongo;
@@ -201,7 +204,7 @@ export class MongoSyncBucketStorage implements SyncRulesBucketStorage {
     checkpoint: util.OpId,
     dataBuckets: Map<string, string>,
     options?: BucketDataBatchOptions
-  ): AsyncIterable<util.SyncBucketData> {
+  ): AsyncIterable<SyncBucketDataBatch> {
     if (dataBuckets.size == 0) {
       return;
     }
@@ -267,6 +270,7 @@ export class MongoSyncBucketStorage implements SyncRulesBucketStorage {
 
     let batchSize = 0;
     let currentBatch: util.SyncBucketData | null = null;
+    let targetOp: bigint | null = null;
 
     // Ordered by _id, meaning buckets are grouped together
     for (let rawData of data) {
@@ -284,7 +288,8 @@ export class MongoSyncBucketStorage implements SyncRulesBucketStorage {
           start = currentBatch.after;
           currentBatch = null;
           batchSize = 0;
-          yield yieldBatch;
+          yield { batch: yieldBatch, targetOp: targetOp };
+          targetOp = null;
         }
 
         start ??= dataBuckets.get(bucket);
@@ -298,17 +303,18 @@ export class MongoSyncBucketStorage implements SyncRulesBucketStorage {
           data: [],
           next_after: start
         };
+        targetOp = null;
+      }
+
+      const entry = mapOpEntry(row);
+
+      if (row.target_op != null) {
+        // MOVE, CLEAR
+        if (targetOp == null || row.target_op > targetOp) {
+          targetOp = row.target_op;
+        }
       }
 
-      const entry: util.OplogEntry = {
-        op_id: util.timestampToOpId(row._id.o),
-        op: row.op,
-        object_type: row.table,
-        object_id: row.row_id,
-        checksum: Number(row.checksum),
-        subkey: `${row.source_table}/${row.source_key.toHexString()}`,
-        data: row.data
-      };
       currentBatch.data.push(entry);
       currentBatch.next_after = entry.op_id;
 
@@ -318,7 +324,8 @@ export class MongoSyncBucketStorage implements SyncRulesBucketStorage {
     if (currentBatch != null) {
       const yieldBatch = currentBatch;
       currentBatch = null;
-      yield yieldBatch;
+      yield { batch: yieldBatch, targetOp: targetOp };
+      targetOp = null;
     }
   }
 
@@ -530,4 +537,8 @@ export class MongoSyncBucketStorage implements SyncRulesBucketStorage {
       }
     );
   }
+
+  async compact(options?: CompactOptions) {
+    return new MongoCompactor(this.db, this.group_id, options).compact();
+  }
 }

package/src/storage/mongo/models.ts

@@ -42,12 +42,13 @@ export interface BucketParameterDocument {
 export interface BucketDataDocument {
   _id: BucketDataKey;
   op: OpType;
-  source_table: bson.ObjectId;
-  source_key: bson.UUID;
-  table: string;
-  row_id: string;
+  source_table?: bson.ObjectId;
+  source_key?: bson.UUID;
+  table?: string;
+  row_id?: string;
   checksum: number;
   data: string | null;
+  target_op?: bigint | null;
 }
 
 export type OpType = 'PUT' | 'REMOVE' | 'MOVE' | 'CLEAR';

package/src/storage/mongo/util.ts

@@ -2,6 +2,9 @@ import { SqliteJsonValue } from '@powersync/service-sync-rules';
 import * as bson from 'bson';
 import * as mongo from 'mongodb';
 import * as crypto from 'crypto';
+import { BucketDataDocument } from './models.js';
+import { timestampToOpId } from '../../util/utils.js';
+import { OplogEntry } from '../../util/protocol-types.js';
 
 /**
  * Lookup serialization must be number-agnostic. I.e. normalize numbers, instead of preserving numbers.
@@ -86,3 +89,25 @@ export const BSON_DESERIALIZE_OPTIONS: bson.DeserializeOptions = {
   // use bigint instead of Long
   useBigInt64: true
 };
+
+export function mapOpEntry(row: BucketDataDocument): OplogEntry {
+  if (row.op == 'PUT' || row.op == 'REMOVE') {
+    return {
+      op_id: timestampToOpId(row._id.o),
+      op: row.op,
+      object_type: row.table,
+      object_id: row.row_id,
+      checksum: Number(row.checksum),
+      subkey: `${row.source_table}/${row.source_key!.toHexString()}`,
+      data: row.data
+    };
+  } else {
+    // MOVE, CLEAR
+
+    return {
+      op_id: timestampToOpId(row._id.o),
+      op: row.op,
+      checksum: Number(row.checksum)
+    };
+  }
+}

package/src/sync/RequestTracker.ts

@@ -1,4 +1,3 @@
-import { container } from '@powersync/lib-services-framework';
 import { Metrics } from '../metrics/Metrics.js';
 
 /**
@@ -10,12 +9,13 @@ export class RequestTracker {
 
   addOperationsSynced(operations: number) {
     this.operationsSynced += operations;
-    container.getImplementation(Metrics).operations_synced_total.add(operations);
+
+    Metrics.getInstance().operations_synced_total.add(operations);
   }
 
   addDataSynced(bytes: number) {
     this.dataSyncedBytes += bytes;
 
-    container.getImplementation(Metrics).data_synced_bytes.add(bytes);
+    Metrics.getInstance().data_synced_bytes.add(bytes);
   }
 }

package/src/sync/sync-index.ts

@@ -1,6 +1,7 @@
 export * from './BroadcastIterable.js';
 export * from './LastValueSink.js';
 export * from './merge.js';
+export * from './RequestTracker.js';
 export * from './safeRace.js';
 export * from './sync.js';
 export * from './util.js';

package/src/sync/sync.ts

@@ -8,6 +8,7 @@ import * as storage from '../storage/storage-index.js';
 import * as util from '../util/util-index.js';
 
 import { logger } from '@powersync/lib-services-framework';
+import { Metrics } from '../metrics/Metrics.js';
 import { mergeAsyncIterables } from './merge.js';
 import { TokenStreamOptions, tokenStream } from './util.js';
 import { RequestTracker } from './RequestTracker.js';
@@ -257,6 +258,9 @@ interface BucketDataBatchResult {
 async function* bucketDataBatch(request: BucketDataRequest): AsyncGenerator<BucketDataBatchResult, void> {
   const { storage, checkpoint, bucketsToFetch, dataBuckets, raw_data, binary_data, tracker, signal } = request;
 
+  const checkpointOp = BigInt(checkpoint);
+  let checkpointInvalidated = false;
+
   const [_, release] = await syncSemaphore.acquire();
   try {
     // Optimization: Only fetch buckets for which the checksums have changed since the last checkpoint
@@ -266,13 +270,16 @@ async function* bucketDataBatch(request: BucketDataRequest): AsyncGenerator<Buck
 
     let has_more = false;
 
-    for await (let r of data) {
+    for await (let { batch: r, targetOp } of data) {
       if (signal.aborted) {
         return;
       }
       if (r.has_more) {
         has_more = true;
       }
+      if (targetOp != null && targetOp > checkpointOp) {
+        checkpointInvalidated = true;
+      }
       if (r.data.length == 0) {
         continue;
       }
@@ -308,12 +315,19 @@ async function* bucketDataBatch(request: BucketDataRequest): AsyncGenerator<Buck
     }
 
     if (!has_more) {
-      const line: util.StreamingSyncCheckpointComplete = {
-        checkpoint_complete: {
-          last_op_id: checkpoint
-        }
-      };
-      yield { data: line, done: true };
+      if (checkpointInvalidated) {
+        // Checkpoint invalidated by a CLEAR or MOVE op.
+        // Don't send the checkpoint_complete line in this case.
+        // More data should be available immediately for a new checkpoint.
+        yield { data: null, done: true };
+      } else {
+        const line: util.StreamingSyncCheckpointComplete = {
+          checkpoint_complete: {
+            last_op_id: checkpoint
+          }
+        };
+        yield { data: line, done: true };
+      }
     }
   } finally {
     release();

package/src/sync/util.ts

@@ -1,6 +1,7 @@
 import * as timers from 'timers/promises';
 
 import * as util from '../util/util-index.js';
+import { Metrics } from '../metrics/Metrics.js';
 import { RequestTracker } from './RequestTracker.js';
 
 export type TokenStreamOptions = {

package/src/util/config/collectors/config-collector.ts

@@ -1,6 +1,8 @@
+import * as t from 'ts-codec';
 import * as yaml from 'yaml';
 
 import { configFile } from '@powersync/service-types';
+import { schema } from '@powersync/lib-services-framework';
 
 import { RunnerConfig } from '../types.js';
 
@@ -21,6 +23,13 @@ export enum ConfigFileFormat {
  */
 const YAML_ENV_PREFIX = 'PS_';
 
+// ts-codec itself doesn't give great validation errors, so we use json schema for that
+const configSchemaValidator = schema
+  .parseJSONSchema(
+    t.generateJSONSchema(configFile.powerSyncConfig, { allowAdditional: true, parsers: [configFile.portParser] })
+  )
+  .validator();
+
 export abstract class ConfigCollector {
   abstract get name(): string;
 
@@ -30,6 +39,45 @@ export abstract class ConfigCollector {
    */
   abstract collectSerialized(runnerConfig: RunnerConfig): Promise<configFile.SerializedPowerSyncConfig | null>;
 
+  /**
+   * Collects the PowerSyncConfig settings.
+   * Validates and decodes the config.
+   * @returns null if this collector cannot provide a config
+   */
+  async collect(runner_config: RunnerConfig): Promise<configFile.PowerSyncConfig | null> {
+    const serialized = await this.collectSerialized(runner_config);
+    if (!serialized) {
+      return null;
+    }
+
+    /**
+     * After this point a serialized config has been found. Any failures to decode or validate
+     * will result in a hard stop.
+     */
+    const decoded = this.decode(serialized);
+    this.validate(decoded);
+    return decoded;
+  }
+
+  /**
+   * Validates input config
+   * ts-codec itself doesn't give great validation errors, so we use json schema for that
+   */
+  validate(config: configFile.PowerSyncConfig) {
+    const valid = configSchemaValidator.validate(config);
+    if (!valid.valid) {
+      throw new Error(`Failed to validate PowerSync config: ${valid.errors.join(', ')}`);
+    }
+  }
+
+  decode(encoded: configFile.SerializedPowerSyncConfig): configFile.PowerSyncConfig {
+    try {
+      return configFile.powerSyncConfig.decode(encoded);
+    } catch (ex) {
+      throw new Error(`Failed to decode PowerSync config: ${ex}`);
+    }
+  }
+
   protected parseContent(content: string, contentType?: ConfigFileFormat) {
     switch (contentType) {
       case ConfigFileFormat.YAML: