🪢 Open source agent evals & observability: Trace, evaluate, and improve LLM applications with one open platform.
TypeScript analyticsautogenevaluationlangchainlarge-language-modelsllama-indexllmllm-evaluationllm-observabilityllmopsmonitoringobservabilityopen-sourceopenaiplaygroundprompt-engineeringprompt-managementself-hostedycombinator
import type { Queue } from "bullmq";
import { logger } from "../logger";
// DO NOT REMOVE THIS FILE AND BEHAVIOUR WITHIN A MINOR LANGFUSE VERSION.
// THE CLEANUP BELOW DELETES THE OLD md5-KEYED REPEAT SCHEDULES; WITHOUT IT,
// SELF-HOSTERS UPGRADING FROM A LEGACY-SCHEDULING VERSION GET DUPLICATE CRON
// TRIGGERS FOREVER (bullmq v6 workers keep iterating stranded md5-keyed
// chains; only manual Redis cleanup stops them).
// HOLD THE BULLMQ v6 UPGRADE UNTIL LANGFUSE V5: v6 deletes
// Queue.getRepeatableJobs() and Queue.removeRepeatableByKey(), which this
// cleanup requires.
/**
* Registers a recurring cron job on a queue via a BullMQ job scheduler.
*
* Job schedulers replace the deprecated `Queue.add(name, data, { repeat })`
* API. The legacy path replaced the pending next iteration in two separate
* Redis round trips on every boot (delete the delayed job, then re-create
* it), so a producer failing between the two - e.g. a container killed
* mid-boot during a fleet-wide deploy - silently ended the schedule until
* the next boot re-registered it. `upsertJobScheduler` performs the same
* replacement in a single atomic Lua script, so concurrent boots and
* mid-boot crashes cannot lose the chain. BullMQ v6 removes the legacy API
* entirely.
*
* The scheduler id is the job name. Legacy schedules are keyed by
* md5(name:jobId:endDate:tz:pattern) instead, so this helper also removes the
* legacy entries for the current pattern and any historical ones - without
* that cleanup the old chain keeps firing in parallel with the scheduler. The
* legacy keys are read back from Redis rather than recomputed, because
* recomputing them requires md5 and that is unavailable when OpenSSL runs in
* FIPS mode. The cleanup is idempotent and must stay in place for as long as a
* direct upgrade from a Langfuse version that scheduled via the legacy API is
* supported.
*
* Never rejects: scheduling runs fire-and-forget from queue singleton
* getters, so all failures are logged instead of thrown.
*/
export const scheduleRecurringJob = async (
queue: Queue,
opts: {
jobName: string;
pattern: string;
/**
* Cron patterns this job was registered with via the legacy repeat API
* in earlier releases. The current pattern is always cleaned up and does
* not need to be listed. When changing `pattern`, append the old value
* here, since each legacy pattern was a separate schedule.
*/
previousPatterns?: string[];
/** Template data stored on the scheduler and passed to every iteration. */
data?: Record<string, unknown>;
},
): Promise<void> => {
const { jobName, pattern, previousPatterns = [], data } = opts;
const legacyPatterns = new Set([pattern, ...previousPatterns]);
try {
// Read the stored keys back rather than recomputing them.
// Queue.removeRepeatable() derives md5(name:jobId:endDate:tz:pattern) to
// locate the entry, and an OpenSSL FIPS provider refuses md5 outright, so
// on a FIPS image that call throws ERR_OSSL_EVP_UNSUPPORTED and the
// cleanup silently never happens - producing exactly the duplicate cron
// triggers this file exists to prevent. Enumerating hashes nothing, and
// removal by stored key is hash-free in the Lua script.
// eslint-disable-next-line @typescript-eslint/no-deprecated -- legacy repeat entries are invisible to getJobSchedulers; keep while direct upgrades from legacy-scheduling releases are supported.
const stored = await queue.getRepeatableJobs();
const staleKeys = stored
.filter(
(job) =>
job.name === jobName &&
// The scheduler keeps its own entry in this same sorted set, keyed
// by the scheduler id, which is the job name. Legacy entries are
// keyed by a hash, so an entry keyed by the job name is the live
// scheduler and must not be removed here.
job.key !== jobName &&
job.pattern !== null &&
legacyPatterns.has(job.pattern),
)
.map((job) => job.key);
// One failing key must not leave the remaining legacy chains firing.
for (const key of staleKeys) {
try {
// eslint-disable-next-line @typescript-eslint/no-deprecated -- pairs with getRepeatableJobs above; removeJobScheduler cannot address legacy entries.
await queue.removeRepeatableByKey(key);
} catch (err) {
logger.error(
`Error removing legacy ${jobName} schedule ${key} on ${queue.name}`,
err,
);
}
}
} catch (err) {
logger.error(
`Error listing legacy ${jobName} schedules on ${queue.name}`,
err,
);
}
try {
await queue.upsertJobScheduler(
jobName,
{ pattern },
{ name: jobName, data: data ?? {} },
);
} catch (err) {
logger.error(
`Error upserting ${jobName} scheduler on ${queue.name} - the recurring job may not run until the next boot`,
err,
);
}
};