Root Cause Analysis – Encoding Service Outage 23 Nov 2025

Summary
On 23 Nov 2025, the Bitmovin Encoding Platform experienced a service outage affecting our VoD and Live encoding pipeline. The issue resulted in temporary unavailability of encoding operations including starting and stopping of encodings and unavailability of status updates of running encodings for the time of the outage.

Root Cause

A bug in one of our encoding services caused memory usage and database (DB) data transfer volumes to increase slowly but steadily over the course of approximately one month.

This resulted in:

• Gradually escalating DB read traffic from the encoding services due to increasing object size with time

• Overwhelming the service with growing message queues for this specific service due to decreased throughput

• Data transfer reaching ~900 MB/s outbound from the database to the service instances

As memory usage kept rising, multiple encoding service instances ultimately crashed simultaneously, leading to an abrupt stop in processing and further service degradation due to accumulation of ~300,000 unprocessed service messages during the outage from 18:00 CET.

Impact

• Encoding jobs failed to start after queueing 

• Running encoding jobs failed to update their status to “In Progress” or “Finished” during the incident

Detection

• 18:06 CET – Investigation initiated based on internal monitoring and alerts.

• 18:07 CET – Outage confirmed and posted on status page.

Internally, teams observed service pods crashing due to memory limit exhaustion, causing message processing interruptions and queue growth.

Investigation, Mitigation & Recovery

After the service crashes, we initiated a controlled process:

1. Increased service memory limits on our kubernetes cluster 18:25 CET

2. Reduced parallel message queue worker logic of the affected service to reduce DB read load by about 30% 19:02 CET

3. Implemented further monitoring and tracing in the service to increase visibility into the massive DB data streaming causes 19:38 CET

4. Re-routed several messages on our messaging system to further decrease DB read load 19:51 CET

5. Sped up processing of messages causing excessive read operations through service based skipping with a new service version 20:23 CET

6. Deployed new service which decreased DB read load to normal levels 20:56

7. Increased message throughput for services to restore the system state by processing the message backlog of about 350k messages in RabbitMQ 21:30 CET

8. Restored all service configuration and gradually increased encoding throughput 23:00 CET

9. Full service restored 23:31 CET

Next Steps and Preventive Actions

To prevent recurrence, we are implementing the following:

• Implement a holistic fix for the bug causing excessive DB read load

• Expand message queue and memory usage monitoring with stricter alerts for read amplification

• Comprehensive review of encoding service code paths related to heavy DB reads

• Message queue pressure safeguards including throttling for struggling services

Our systems are stable, and we are no longer observing any errors or delays.

A Root Cause Analysis (RCA) will follow once our internal review is completed. Thank you for your patience throughout this disruption.

We are increasing parallel encoding slots every 10 minutes and the queued encoding backlog built up in during the incident will be finished soon. We encourage our customers to restart potentially stopped workflows again.

Our internal system state has recovered and we are slowly increasing parallel processing throughput for the queued encoding backlog.

We have identified the root cause of the issue and implemented a fix, encoding backlog processing has sped up but will take more time to fully recover the overall system state.

Our team has likely identified the root cause of high DB load causing delays in encoding status updates in our API and dashboard. We are deploying a new version of the affected service and will have an update in a few minutes.

Our team is continuing to investigate the issue impacting slow API requests, delayed encoding status updates and delayed manifest generations (manifests generated with the encoding start calls). Encoding output is not affected.

At this time, we have not yet identified the root cause. Diagnostic efforts are ongoing, and all necessary teams are engaged.

We sincerely apologize for the inconvenience and will provide updates as soon as we have more information.

Thank you for your patience and understanding.

We have successfully reduced the Encoding API error rate; however, our services are still struggling to process the backlog of queued service messages.

To help stabilise the system, we have further lowered the VOD processing limits and continue to actively work on clearing the queue and restoring full service performance.

Webhook notifications are also lagging behind.

Further updates will follow as we make progress.

We are currently experiencing a partial outage of the Encoding API. Error rates remain elevated and queued encodings continue to accumulate.

To stabilise the platform, we have temporarily reduced maximum processing concurrency. While this measure helps improve system stability, it also impacts throughput and leads to longer wait times for queued jobs.

Our engineering team is actively working on restoring normal performance. We will provide further updates as we make progress.

Incident Post-Mortem – Encoding Cleanup Gap-Lock

Date: 2025-10-16
Duration: 07:49–08:30 UTC (41 minutes)
Customer Impact: Calls to the https://api.bitmovin.com/v1/encoding/encodings/{encoding_id}/muxings/ endpoints failed, leading to workflow stoppages. The failure rate remained below monitoring thresholds, so the issue was not immediately detected.

Summary

At 07:49 UTC, our encoding cleanup process (responsible for deleting old encodings outside of retention) triggered a gap-lock on one of our database tables. While the long-running delete transaction was active, the API was unable to insert new records into the table.

This directly impacted the muxings endpoints, which failed during this period and caused encoding workflows to stop. Because the error rate was below our monitoring alert thresholds, the incident went unnoticed until the team was alerted at 08:20 UTC.

Root Cause

The issue was caused by outdated MySQL table statistics, which led the MySQL query optimizer to select an inefficient execution plan:

• Instead of using an index scan on the large table, the query optimizer chose a full table scan.

• Under REPEATABLE_READ isolation, this resulted in gap-locks across wide portions of the table.

• In practice, this behaved like a table-level lock, blocking all inserts until the cleanup query was stopped.

Timeline

• 07:49 UTC – Encoding cleanup process starts. Optimizer chooses full table scan → gap-locks prevent inserts.

• 07:49–08:20 UTC – Calls to /encoding/encodings/{encoding_id}/muxings/ fail, workflows stop. Monitoring does not alert as errors are below configured threshold.

• 08:20 UTC – Engineering team is alerted.

• 08:20–08:30 UTC – Cleanup process identified as root cause. Process is stopped, lock released.

• 08:30 UTC – Incident resolved, API resumes normal operation.

Immediate Actions

After resolving the incident, we immediately stopped the cleanup process and investigated why it suddenly caused issues despite having run successfully for years.

Since then, we have:

• Tightened timeouts for deletion commands in the cleanup process.

• Changed to a more forgiving isolation level to prevent broad blocking locks.

• Adjusted the cleanup process to use smaller batch sizes and higher concurrency, which reduced database impact and increased throughput.

• Added significantly more monitoring to the cleanup process to better track query performance and catch anomalies early.

Next Steps

• Update MySQL statistics so the query optimizer selects the correct execution plan.

  • Expected completion: by end of this week (2025-10-24)

• Rework alerting strategy to detect low-rate but workflow-blocking errors earlier.

  • Expected completion: within 3 weeks (by 2025-11-06)

We experienced elevated API error rates on the encoding endpoints between 07:50 and 08:30 UTC. The issue, caused by the encoding service during muxing creation, was identified and promptly resolved. The engineering team is preparing a post-mortem and will share it here once available.