Candidate B production availability incident

Customer impact

On 17–18 August, reads and writes through mdbase hosted storage, Connect, and MCP were unavailable during two UTC windows: 23:37:04–23:45:17 on 17 August, and 23:51:38 on 17 August–04:15:16 on 18 August. Requests returned 503 responses. The first window is conservatively measured through the completed restoration, and the second ends when normal authenticated reads and writes resumed.

The release also caused two narrower hosted-storage interruptions on 18 August: approximately 09:2x–09:44 UTC (about 20 minutes), and 11:49:00–11:49:33 UTC. Hosted collection operations returned 502 responses; Connect and MCP remained available during the second interruption.

Data integrity

No canonical encrypted records were lost or rewritten. Exact Markdown remained the authoritative copy throughout recovery. Projection integrity was verified after the upgrade, and authenticated record and file write-read-delete checks passed after service restoration. There was no identified confidentiality impact.

Correction

The rollout was completed using durable progress checkpoints and additional verification before traffic reopened. Deployment checks were corrected so a closed service is tested for liveness rather than readiness. The later hosted-storage defect was fixed by keeping unaffected projections current when a saved view changes and by preventing one degraded collection from withdrawing the whole provider. Future storage migrations require measured production-shape rehearsals, a bounded outage decision, resumable recovery testing, and independent post-deployment read/write checks.

Beta service

Connect remains a beta service with no availability SLA. This permanent summary records how a beta failure affected customers and what changed; it is not an uptime commitment. Current conditions are published on the service status page.