v2.15.0
🗜️ Notification Compression
Oban uses notifications across most core functionality, from job staging to cancellation. Some notifications, such as gossip, contain massive redundancy that compresses nicely. For example, this table breaks down the compression ratios for a fairly standard gossip payload containing data from ten queues:
Mode | Bytes | % of Original |
---|---|---|
Original | 4720 | 100% |
Gzip | 307 | 7% |
Encode 64 | 412 | 9% |
Minimizing notification payloads is especially important for Postgres because it applies an 8kb limit to all messages. Now all pub/sub notifications are compressed automatically, with a safety mechanism for compatibility with external notifiers, namely Postgres triggers.
🗃️ Query Improvements
There has been an ongoing issue with systems recording a job attempt twice, when it only executed once. While that sounds minor, it could break an entire queue when the attempt exceeded max attempts because it would violate a database constraint.
Apparently, the Postgres planner may choose to generate a plan that executes a nested loop over the LIMITing subquery, causing more UPDATEs than LIMIT. That could cause unexpected updates, including attempts > max_attempts in some cases. The solution is to use a CTE as an "optimization fence" that forces Postgres not to optimize the query.
We also worked in a few additional query improvements:
- Use an index only scan for job staging to safely handle tables with millions of scheduled jobs.
- Remove unnecessary row locking from staging and pruning queries.
🪶 New Engine Callbacks for SQL Compatibility
We're pleased to share improvements in Oban's SQLite integration. A few SQLite pioneers identified pruning and staging compatibility bugs, and instead of simply patching around the issues with conditional logic, we tackled them with new engine callbacks: stage_jobs/3
and prune_jobs/3
. The result is safer, optimized queries for each specific database.
Introducing new engine callbacks with database-specific queries paves the way for working with other databases. There's even an open issue for MySQL support...
v2.15.0 — 2023-04-13
Enhancements
-
[Oban] Use DynamicSupervisor to supervise queues for optimal shutdown
Standard supervisors shut down in a fixed order, which could make shutting down queues with active jobs and a lengthy grace period very slow. This switches to a
DynamicSupervisor
for queue supervision so queues can shut down simultaneously while still respecting the grace period. -
[Executor] Retry acking infinitely after job execution
After jobs execute the producer must record their status in the database. Previously, if acking failed due to a connection error after 10 retries it would orphan the job. Now, acking retries infinitely (with backoff) until the function succeeds. The result is stronger execution guarantees with backpressure during periods of database fragility.
-
[Oban] Accept a
Job
struct as well as a job id forcancel_job/1
andretry_job/1
Now it's possible to write
Oban.cancel_job(job)
directly rather thanOban.cancel_job(job.id)
. -
[Worker] Allow snoozing jobs for zero seconds.
Returning
{:snooze, 0}
immediately reschedules a job without any delay. -
[Notifier] Accept arbitrary channel names for notifications, e.g. "my-channel"
-
[Telemetry] Add 'detach_default_logger/0' to programmatically disable an attached logger.
-
[Testing] Avoid unnecessary query for "happy path" assertion errors in
assert_enqueued/2
-
[Testing] Inspect charlists as lists in testing assertions
Args frequently contain lists of integers like
[123]
, which was curiously displayed as'{'
.
Bug Fixes
-
[Executor] Correctly raise "unknown worker" errors.
Unknown workers triggered an unknown case error rather than the appropriate "unknown worker" runtime error.
-
[Testing] Allow
assert_enqueued
with ascheduled_at
time foravailable
jobsThe use of
Job.new
to normalize query fields would change assertions with a "scheduled_at" date to only check scheduled, never "available" -
[Telemetry] Remove
:worker
from engine and plugin query meta.The
worker
isn't part of any query indexes and prevents optimal index usage. -
[Job] Correct priority type to cover default of 0
For changes prior to v2.15 see the v2.14 docs.