Resolved incident

Python Infrastructure Outage

PyPI · Critical · Started

Duration
2h 47m
Severity
Critical
Status
Postmortem

Summary

Critical · Postmortem

The incident has been resolved. ## Incident Report: Cluster-wide service outage **Duration**: ~36 minutes \(20:04–20:40 UTC\) **Impact**: Some PSF-hosted services were unavailable, including [python.org](http://python.org), [us.pycon.org](http://us.pycon.org), PyPI stats, [bugs.python.org](http://bugs.python.org), and related services. What was unaffected was our other cluster that manages [PyPI.org](http://PyPI.org) among other services related to PyPI. **Root Cause**: During local development of kubernetes workloads locally there was an incorrect context switch to one of our production clusters. The scale-down commands ran against the production cluster instead of the local environment, iterating through all deployments and setting them to zero replicas. which created cascading failures. **Recovery**: Services were restored with the help of Ee Durbin by bringing up infrastructure in dependency order, original replica counts were recovered from Kubernetes event history. Action items: * Separate kubeconfig files for production vs local, rather than relying on context switching * Research adding admission control or policies to prevent bulk scale-to-zero operations * Document the infrastructure dependency chain and recovery runbook for future incidents Jacob Coffee, PSF Infrastructure Team

View official incident ↗

Event timeline

How this incident unfolded, from the official status page.

  1. Investigating

    PyPI

    We are currently investigating this issue

  2. Investigating

    PyPI

    The root cause has been identified as an erroneous scaling operation that affected some application workloads. All services have been restored and are coming back online. We are monitoring to confirm full recovery.

  3. Resolved

    PyPI

    The incident has been resolved.

  4. Postmortem

    PyPI

    The incident has been resolved. ## Incident Report: Cluster-wide service outage **Duration**: ~36 minutes \(20:04–20:40 UTC\) **Impact**: Some PSF-hosted services were unavailable, including [python.org](http://python.org), [us.pycon.org](http://us.pycon.org), PyPI stats, [bugs.python.org](http://bugs.python.org), and related services. What was unaffected was our other cluster that manages [PyPI.org](http://PyPI.org) among other services related to PyPI. **Root Cause**: During local development of kubernetes workloads locally there was an incorrect context switch to one of our production clusters. The scale-down commands ran against the production cluster instead of the local environment, iterating through all deployments and setting them to zero replicas. which created cascading failures. **Recovery**: Services were restored with the help of Ee Durbin by bringing up infrastructure in dependency order, original replica counts were recovered from Kubernetes event history. Action items: * Separate kubeconfig files for production vs local, rather than relying on context switching * Research adding admission control or policies to prevent bulk scale-to-zero operations * Document the infrastructure dependency chain and recovery runbook for future incidents Jacob Coffee, PSF Infrastructure Team