Operations and performance
The demo server and private hosted MCP service provide integration references. Static grants and single-host recovery do not provide a complete production deployment.
| Area | What to plan for |
|---|---|
| Access | The private hosted MCP service has static bearer grants. Application document-server authorization, transport security, and tenant isolation remain deployment responsibilities. |
| Storage | Select and operate durable storage, backups, restore procedures, and retention policies. Test failure modes on the actual storage system. |
| Recovery | The MCP journal supports macOS and Linux on private local storage; Linux requires /usr/bin/flock. Each journal record is limited to 16 MiB. Use a stable, separate journal directory for each participant. Protect its document contents. |
| Lifecycle | Monitor pending edits, recovery reservations, garbage collection, reconnects, and unresolved outcomes. Never interpret an unknown receipt as rejection. |
| Capacity | Measure with your own document sizes, edit history, concurrent clients, network, editor, and hardware. Include browser rendering and memory, not only core operations. |
See Troubleshooting and the recovery plan before operating a durable session.
Example measurement: long edit history
A local demo-server benchmark used a 5,093,156-byte log with 30,000 CRDT changes and one visible character. It measures log replay and separate document I/O, not rendering a large document. On an Apple M3 Max with Node 24, two warmups and 12 recorded samples per version gave these HTTP response medians:
| Request | Before | After |
|---|---|---|
| First state request for the cold log | 258.56 ms | 263.52 ms |
| Append to a different warm document during replay | 263.31 ms | 13.94 ms |
The cold request did not get faster in this run. The host was shared, runs were sequential, and the 12-sample nearest-rank p95 is the maximum. These numbers do not establish production latency or browser responsiveness. The full report records the workload, checks, source hashes, and limitations. See Performance evidence for browser and core measurements and their separate scopes.