@dbos-inc/dbos-sdk 4.26.10 → 4.27.3-preview

package/dist/src/system_database.js

@@ -15,6 +15,7 @@ const pg_1 = require("pg");
 const error_1 = require("./error");
 const workflow_1 = require("./workflow");
 const utils_1 = require("./utils");
+const wfqueue_1 = require("./wfqueue");
 const crypto_1 = require("crypto");
 const utils_2 = require("./utils");
 const database_utils_1 = require("./database_utils");
@@ -47,9 +48,15 @@ const QUEUE_COLUMN_BY_FIELD = {
     rateLimitPeriodSec: 'rate_limit_period_sec',
     priorityEnabled: 'priority_enabled',
     partitionQueue: 'partition_queue',
+    partitionConcurrency: 'partition_concurrency',
+    partitionWorkerConcurrency: 'partition_worker_concurrency',
+    partitionRateLimitMax: 'partition_rate_limit_max',
+    partitionRateLimitPeriodSec: 'partition_rate_limit_period_sec',
     pollingIntervalSec: 'polling_interval_sec',
 };
-const QUEUE_COLUMNS = 'name, concurrency, worker_concurrency, rate_limit_max, rate_limit_period_sec, priority_enabled, partition_queue, polling_interval_sec, application_name';
+const QUEUE_COLUMNS = 'name, concurrency, worker_concurrency, rate_limit_max, rate_limit_period_sec, priority_enabled, partition_queue, ' +
+    'partition_concurrency, partition_worker_concurrency, partition_rate_limit_max, partition_rate_limit_period_sec, ' +
+    'polling_interval_sec, application_name';
 function queueRecordFromRow(row) {
     return {
         name: row.name,
@@ -59,6 +66,10 @@ function queueRecordFromRow(row) {
         rateLimitPeriodSec: row.rate_limit_period_sec,
         priorityEnabled: row.priority_enabled,
         partitionQueue: row.partition_queue,
+        partitionConcurrency: row.partition_concurrency,
+        partitionWorkerConcurrency: row.partition_worker_concurrency,
+        partitionRateLimitMax: row.partition_rate_limit_max,
+        partitionRateLimitPeriodSec: row.partition_rate_limit_period_sec,
         pollingIntervalSec: row.polling_interval_sec,
         applicationName: row.application_name ?? undefined,
     };
@@ -1635,7 +1646,17 @@ class SystemDatabase {
     clearRunningWorkflow(workflowID) {
         this.runningWorkflowMap.delete(workflowID);
     }
-    countRunningWorkflowsForQueue(queueName, queuePartitionKey) {
+    /** Workflows this worker is running for a queue, across every partition of it. */
+    countRunningWorkflowsForQueue(queueName) {
+        let count = 0;
+        for (const entry of this.runningWorkflowMap.values()) {
+            if (entry.queueName === queueName)
+                count++;
+        }
+        return count;
+    }
+    /** Workflows this worker is running for one partition of a queue. */
+    countRunningWorkflowsForPartition(queueName, queuePartitionKey) {
         let count = 0;
         for (const entry of this.runningWorkflowMap.values()) {
             if (entry.queueName === queueName && entry.queuePartitionKey === queuePartitionKey)
@@ -2334,78 +2355,93 @@ class SystemDatabase {
        SELECT pk FROM partitions WHERE pk IS NOT NULL`, params);
         return rows.map((row) => row.pk);
     }
-    async findAndMarkStartableWorkflows(queue, executorID, appVersion, queuePartitionKey) {
-        const limiterPeriodMS = queue.rateLimit ? queue.rateLimit.periodSec * 1000 : 0;
+    async findAndMarkStartableWorkflows(queue, executorID, appVersion, queuePartitionKey, localRunningCount = 0, partitionLocalRunningCount = 0) {
         const claimedIDs = [];
-        const localRunningForQueue = this.countRunningWorkflowsForQueue(queue.name, queuePartitionKey);
-        // Build partition key filter
-        let partitionFilter = '';
-        const partitionParams = [];
-        if (queuePartitionKey !== undefined) {
-            partitionFilter = `AND queue_partition_key = $PARTITION`;
-            partitionParams.push(queuePartitionKey);
-        }
+        const limits = (0, wfqueue_1.resolveQueueLimits)(queue);
+        const partitionParams = queuePartitionKey !== undefined ? [queuePartitionKey] : [];
+        // Shares a concurrency or rate limit budget with other executors.
+        const hasSharedBudget = limits.globalConcurrency !== undefined ||
+            limits.partitionConcurrency !== undefined ||
+            limits.rateLimit !== undefined ||
+            limits.partitionRateLimit !== undefined;
+        // Shares that budget across partitions too, so sweeps of different partitions read disjoint rows and could each spend it.
+        const hasWriteSkew = queuePartitionKey !== undefined && (limits.globalConcurrency !== undefined || limits.rateLimit !== undefined);
         const client = await this.#connect();
         try {
-            // Default to READ COMMITTED except with global concurrency limits or rate limits
-            if (queue.concurrency !== undefined || queue.rateLimit !== undefined) {
-                await client.query('BEGIN ISOLATION LEVEL REPEATABLE READ');
+            // Default to READ COMMITTED except with a budget shared across executors
+            if (hasSharedBudget) {
+                await client.query(`BEGIN ISOLATION LEVEL ${hasWriteSkew ? 'SERIALIZABLE' : 'REPEATABLE READ'}`);
             }
             else {
                 await client.query('BEGIN');
             }
-            // If there is a rate limit, compute how many functions have started in its period.
-            let numRecentQueries = 0;
-            if (queue.rateLimit) {
-                const params = [
-                    queue.name,
-                    workflow_1.StatusString.ENQUEUED,
-                    workflow_1.StatusString.DELAYED,
-                    limiterPeriodMS,
-                    ...partitionParams,
-                ];
+            /** Slots left in a rate limit's rolling window, at the scope that limit applies to. */
+            const rateLimitRemaining = async (rateLimit, partitionScoped) => {
+                const params = [queue.name, workflow_1.StatusString.ENQUEUED, workflow_1.StatusString.DELAYED, rateLimit.periodSec * 1000];
                 // Count only what this application would dequeue, matching the select below.
                 const scope = this.#appNameFilter('application_name', this.appName, params);
-                const countResult = await client.query(`SELECT COUNT(*) FROM "${this.schemaName}".workflow_status
+                const partitionFilter = partitionScoped ? `AND queue_partition_key = $${params.push(queuePartitionKey)}` : '';
+                const { rows } = await client.query(`SELECT COUNT(*) FROM "${this.schemaName}".workflow_status
            WHERE queue_name = $1
              AND rate_limited = TRUE
              AND status NOT IN ($2, $3)
              -- Database clock on both sides, as the claim stamps started_at_epoch_ms with it.
              AND started_at_epoch_ms > (EXTRACT(epoch FROM now()) * 1000)::bigint - $4
              AND ${scope}
-             ${partitionFilter.replace('$PARTITION', '$5')}`, params);
-                numRecentQueries = Number(countResult.rows[0].count);
-                if (numRecentQueries >= queue.rateLimit.limitPerPeriod) {
-                    await client.query('COMMIT');
-                    return claimedIDs;
-                }
-            }
-            // Dequeue functions eligible for this worker and ordered by the time at which they were enqueued.
-            // If there is a global or local concurrency limit N, select only the N oldest enqueued
-            // functions, else select all of them.
+             ${partitionFilter}`, params);
+                return rateLimit.limitPerPeriod - Number(rows[0].count);
+            };
+            /**
+             * Workflows already running, which peer workers count against too. Kept as its own query per
+             * scope: the partition-scoped predicate rides idx_workflow_status_partition_dequeue_v2, which
+             * a queue-wide scan loses.
+             */
+            const pendingCount = async (partitionScoped) => {
+                const params = [queue.name, workflow_1.StatusString.PENDING];
+                const scope = this.#appNameFilter('application_name', this.appName, params);
+                const partitionFilter = partitionScoped ? `AND queue_partition_key = $${params.push(queuePartitionKey)}` : '';
+                const { rows } = await client.query(`SELECT COUNT(*) FROM "${this.schemaName}".workflow_status
+           WHERE queue_name = $1 AND status = $2 AND ${scope} ${partitionFilter}`, params);
+                return Number(rows[0]?.count ?? 0);
+            };
+            // Compute maxTasks, the number of workflows startable under every flow control limit on this queue.
             let maxTasks = Infinity;
-            if (queue.rateLimit) {
+            if (limits.workerConcurrency !== undefined) {
+                // Use the in-memory registry for this worker's running count — avoids a DB round trip.
+                maxTasks = Math.min(maxTasks, Math.max(0, limits.workerConcurrency - localRunningCount));
+            }
+            if (limits.partitionWorkerConcurrency !== undefined) {
+                maxTasks = Math.min(maxTasks, Math.max(0, limits.partitionWorkerConcurrency - partitionLocalRunningCount));
+            }
+            if (maxTasks <= 0) {
+                await client.query('COMMIT');
+                return claimedIDs;
+            }
+            if (limits.rateLimit !== undefined) {
                 // Bound the claim by the limiter's remaining slots so a backlogged queue locks only what it can start.
-                maxTasks = Math.max(0, queue.rateLimit.limitPerPeriod - numRecentQueries);
+                maxTasks = Math.min(maxTasks, await rateLimitRemaining(limits.rateLimit, false));
             }
-            if (queue.workerConcurrency !== undefined) {
-                // Use the in-memory registry for this worker's running count — avoids a DB round trip.
-                maxTasks = Math.min(maxTasks, Math.max(0, queue.workerConcurrency - localRunningForQueue));
+            if (limits.partitionRateLimit !== undefined) {
+                maxTasks = Math.min(maxTasks, await rateLimitRemaining(limits.partitionRateLimit, true));
             }
-            if (queue.concurrency !== undefined) {
+            if (maxTasks <= 0) {
+                await client.query('COMMIT');
+                return claimedIDs;
+            }
+            if (limits.globalConcurrency !== undefined) {
                 // Global concurrency still requires a DB query since other workers may be running workflows too.
-                const params = [queue.name, workflow_1.StatusString.PENDING, ...partitionParams];
-                const scope = this.#appNameFilter('application_name', this.appName, params);
-                const runningTasksResult = await client.query(`SELECT COUNT(*) as task_count
-           FROM "${this.schemaName}".workflow_status
-           WHERE queue_name = $1 AND status = $2 AND ${scope}
-             ${partitionFilter.replace('$PARTITION', '$3')}`, params);
-                const totalRunningTasks = Number(runningTasksResult.rows[0]?.task_count ?? 0);
-                if (totalRunningTasks > queue.concurrency) {
-                    this.logger.warn(`Total running tasks (${totalRunningTasks}) exceeds the global concurrency limit (${queue.concurrency})`);
+                const totalRunningTasks = await pendingCount(false);
+                if (totalRunningTasks > limits.globalConcurrency) {
+                    this.logger.warn(`Total running tasks (${totalRunningTasks}) exceeds the global concurrency limit (${limits.globalConcurrency})`);
+                }
+                maxTasks = Math.min(maxTasks, Math.max(0, limits.globalConcurrency - totalRunningTasks));
+            }
+            if (limits.partitionConcurrency !== undefined) {
+                const partitionRunningTasks = await pendingCount(true);
+                if (partitionRunningTasks > limits.partitionConcurrency) {
+                    this.logger.warn(`Total running tasks (${partitionRunningTasks}) on partition ${queuePartitionKey} of queue ${queue.name} exceeds the partition concurrency limit (${limits.partitionConcurrency})`);
                 }
-                const availableTasks = Math.max(0, queue.concurrency - totalRunningTasks);
-                maxTasks = Math.min(maxTasks, availableTasks);
+                maxTasks = Math.min(maxTasks, Math.max(0, limits.partitionConcurrency - partitionRunningTasks));
             }
             // Return immediately if there are no available tasks due to flow control limits
             if (maxTasks <= 0) {
@@ -2420,8 +2456,7 @@ class SystemDatabase {
                 : 'application_version = $3';
             // A limit shared across processes needs a consistent view of the table: NOWAIT makes an
             // overlapping dequeuer abort rather than claim the next rows and spend the same budget twice.
-            const sharedBudget = queue.concurrency !== undefined || queue.rateLimit !== undefined;
-            const lockMode = sharedBudget ? 'FOR UPDATE NOWAIT' : 'FOR UPDATE SKIP LOCKED';
+            const lockMode = hasSharedBudget ? 'FOR UPDATE NOWAIT' : 'FOR UPDATE SKIP LOCKED';
             const limitClause = maxTasks !== Infinity ? `LIMIT ${maxTasks}` : '';
             const selectParams = [workflow_1.StatusString.ENQUEUED, queue.name, appVersion, ...partitionParams];
             const selectScope = this.#appNameFilter('application_name', this.appName, selectParams);
@@ -2432,7 +2467,7 @@ class SystemDatabase {
           AND queue_name = $2
           AND ${versionClause}
           AND ${selectScope}
-          ${partitionFilter.replace('$PARTITION', '$4')}
+          ${queuePartitionKey !== undefined ? 'AND queue_partition_key = $4' : ''}
         ORDER BY priority ASC, created_at ASC
         ${limitClause}
         ${lockMode}
@@ -2449,7 +2484,7 @@ class SystemDatabase {
                     workflow_1.StatusString.PENDING,
                     executorID,
                     appVersion,
-                    queue.rateLimit !== undefined,
+                    limits.rateLimit !== undefined || limits.partitionRateLimit !== undefined,
                     workflowIDs,
                     workflow_1.StatusString.ENQUEUED,
                     // Claim an unclaimed row for this application; a nameless dequeuer leaves ownership untouched.
@@ -2492,14 +2527,18 @@ class SystemDatabase {
         // Return the IDs of all functions we marked started
         return claimedIDs;
     }
-    /** Max heads admitted per sweep: bounds dispatch, not the partition walk; lowest keys win, so higher keys can wait under sustained load. */
+    /** Max heads admitted per sweep: bounds dispatch, not the partition walk; an unconstrained sweep takes the lowest keys, so higher keys can wait under sustained load. */
     partitionedDequeueSweepCap = 8192;
-    /** Dequeue each partition's head-of-line workflow in one transaction, at most {@link partitionedDequeueSweepCap} per sweep; only valid for concurrency=1, no-limiter queues. */
-    async findAndMarkStartablePartitionedWorkflows(queue, executorID, appVersion) {
-        if (queue.concurrency !== 1 || queue.rateLimit !== undefined) {
-            throw new error_1.DBOSError(`Batched partitioned dequeue requires a queue with concurrency 1 and no rate limit: ${queue.name}`);
-        }
-        // workerConcurrency needs no handling here: dispatch routes 0 (the pause-dequeue idiom) to the fallback path, and validation caps any other value at concurrency=1, which the PENDING gate already enforces globally.
+    /** Dequeue each partition's head-of-line workflow in one transaction, at most {@link partitionedDequeueSweepCap} per sweep; only valid for partition-concurrency-1 queues with no queue-wide limit. */
+    async findAndMarkStartablePartitionedWorkflows(queue, executorID, appVersion, maxTasks = Infinity) {
+        const limits = (0, wfqueue_1.resolveQueueLimits)(queue);
+        if (limits.partitionConcurrency !== 1 ||
+            limits.globalConcurrency !== undefined ||
+            limits.rateLimit !== undefined ||
+            limits.partitionRateLimit !== undefined) {
+            throw new error_1.DBOSError(`Batched partitioned dequeue requires a queue with partition concurrency 1 and no queue-wide concurrency or rate limit: ${queue.name}`);
+        }
+        // partitionWorkerConcurrency needs no handling here: any value above 0 is capped at partition concurrency 1, which the PENDING gate already enforces globally, and 0 makes the caller's maxTasks 0.
         const client = await this.#connect();
         try {
             await client.query('BEGIN');
@@ -2508,12 +2547,16 @@ class SystemDatabase {
             const versionClause = (n) => isLatestVersion
                 ? `(application_version = $${n} OR application_version IS NULL)`
                 : `application_version = $${n}`;
+            // This worker's own budget bounds the sweep alongside the cap.
+            const sweepLimit = Math.min(this.partitionedDequeueSweepCap, maxTasks);
+            // When the worker's budget is the binding constraint, probe partitions in random order to prevent starvation.
+            const sweepOrder = sweepLimit < this.partitionedDequeueSweepCap ? 'random()' : 'partitions.pk ASC';
             const candidateParams = [
                 queue.name,
                 workflow_1.StatusString.ENQUEUED,
                 appVersion,
                 workflow_1.StatusString.PENDING,
-                this.partitionedDequeueSweepCap,
+                sweepLimit,
             ];
             const candidateScope = this.#appNameFilter('application_name', this.appName, candidateParams);
             // Walk distinct partition keys with a recursive-CTE loose index scan (one seek per key, mirroring getQueuePartitions) so sweep cost scales with partition count, not backlog depth.
@@ -2528,30 +2571,36 @@ class SystemDatabase {
             FROM partitions
             WHERE partitions.pk IS NOT NULL)
          )
+         , chosen AS (
+           SELECT partitions.pk
+           FROM partitions
+           WHERE partitions.pk IS NOT NULL
+             -- Unscoped by design: a mutual-exclusion probe must block on any owner's row.
+             AND NOT EXISTS (
+               SELECT 1
+               FROM "${this.schemaName}".workflow_status
+               WHERE queue_name = $1 AND status = $4
+                 AND queue_partition_key IS NOT NULL AND queue_partition_key = partitions.pk
+             )
+           ORDER BY ${sweepOrder}
+           LIMIT $5
+         )
          SELECT head.workflow_uuid
-         FROM partitions
+         FROM chosen
          -- LATERAL plans as a tight nested loop; a correlated scalar subquery runs as a slower per-row SubPlan.
          JOIN LATERAL (
            SELECT workflow_uuid
            FROM "${this.schemaName}".workflow_status
            WHERE queue_name = $1 AND status = $2
-             AND queue_partition_key = partitions.pk
+             AND queue_partition_key = chosen.pk
              AND ${versionClause(3)}
              AND ${candidateScope}
            -- workflow_uuid totalizes the head order (same head for every worker under created_at ties) and the index's trailing workflow_uuid keeps this a pure top-1 probe.
            ORDER BY priority ASC, created_at ASC, workflow_uuid ASC
            LIMIT 1
          ) head ON TRUE
-         WHERE partitions.pk IS NOT NULL
-           -- Unscoped by design: a mutual-exclusion probe must block on any owner's row.
-           AND NOT EXISTS (
-             SELECT 1
-             FROM "${this.schemaName}".workflow_status
-             WHERE queue_name = $1 AND status = $4
-               AND queue_partition_key IS NOT NULL AND queue_partition_key = partitions.pk
-           )
-         ORDER BY partitions.pk ASC
-         LIMIT $5`, candidateParams);
+         -- Which partitions were chosen is settled above; order the result so the claim, and the dispatch it feeds, are deterministic.
+         ORDER BY chosen.pk ASC`, candidateParams);
             const candidateIDs = candidateResult.rows.map((row) => row.workflow_uuid);
             if (candidateIDs.length === 0) {
                 await client.query('COMMIT');
@@ -3483,6 +3532,10 @@ class SystemDatabase {
           rate_limit_period_sec = EXCLUDED.rate_limit_period_sec,
           priority_enabled = EXCLUDED.priority_enabled,
           partition_queue = EXCLUDED.partition_queue,
+          partition_concurrency = EXCLUDED.partition_concurrency,
+          partition_worker_concurrency = EXCLUDED.partition_worker_concurrency,
+          partition_rate_limit_max = EXCLUDED.partition_rate_limit_max,
+          partition_rate_limit_period_sec = EXCLUDED.partition_rate_limit_period_sec,
           polling_interval_sec = EXCLUDED.polling_interval_sec,
           updated_at = EXCLUDED.updated_at,
           -- Claim only an unclaimed row, so a registration landing between the check above and this write keeps the name it just took.
@@ -3497,8 +3550,10 @@ class SystemDatabase {
             const resolvedOwner = await this.#resolveRowOwner(client, 'queues', 'name', record.name, owner, 'Queue');
             await client.query(`INSERT INTO "${this.schemaName}".queues
           (name, concurrency, worker_concurrency, rate_limit_max, rate_limit_period_sec,
-           priority_enabled, partition_queue, polling_interval_sec, updated_at, application_name)
-         VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9, $10)
+           priority_enabled, partition_queue, partition_concurrency, partition_worker_concurrency,
+           partition_rate_limit_max, partition_rate_limit_period_sec,
+           polling_interval_sec, updated_at, application_name)
+         VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9, $10, $11, $12, $13, $14)
          ${onConflict}`, [
                 record.name,
                 record.concurrency,
@@ -3507,6 +3562,10 @@ class SystemDatabase {
                 record.rateLimitPeriodSec,
                 record.priorityEnabled,
                 record.partitionQueue,
+                record.partitionConcurrency,
+                record.partitionWorkerConcurrency,
+                record.partitionRateLimitMax,
+                record.partitionRateLimitPeriodSec,
                 record.pollingIntervalSec,
                 now,
                 resolvedOwner ?? null,