Sync init times out mid-chunk, /sync/stop times out too, and the watcher stays down (2.11.3–2.12.2)

Back to board
## ▼ FIELD 1 — Steps to reproduce **Environment** | | | |---|---| | MCP server / plugin | most recent incident on **2.12.2**; earlier ones on 2.11.3 and 2.12.1 | | AI client | Claude Code 2.1.227 | | Model | Claude Opus 5 | | OS | Windows 11 Pro (10.0.26200) | | Studio | 0.733.0.7330989 | | Tier | PRO | **Steps** No special setup and no deterministic recipe — this fires on ordinary sync init, on a place that syncs successfully at other times (the same place completed a full sync of 5409 instances / 104 scripts the day before the latest incident). 1. Start Studio on a place whose script payload is on the larger side (ours splits into 39 chunks). 2. Let the sync initialise, either automatically on connect or via the plugin. 3. Observe the plugin log. The incidents are captured in **Roblox Studio's own log files** — that is where the full text survives, since the plugin's in-memory buffer overwrites itself: ```powershell Get-ChildItem "$env:LOCALAPPDATA\Roblox\logs" -Filter *.log | Select-String -Pattern "Failed to send chunk|session cleanup failed" ``` ⚠️ Those logs rotate after roughly 4 days, which is why the older incidents below come from our own notes rather than from quotable log lines. --- ## ▼ FIELD 2 — Expected behavior Sync init either completes, or fails and **recovers**: the client retries, or at minimum the failure is surfaced so the user knows the mirror is no longer being updated. A sync that has silently stopped should not look the same as a sync that is up to date. --- ## ▼ FIELD 3 — Actual behavior **3a — The signature: init times out mid-chunk, and so does the cleanup.** Verbatim from the Studio log, most recent incident (2026-08-11, plugin **2.12.2**): ``` 2026-08-11T07:01:56.328Z Warning [FLog::CreatorWarning] [SyncManager] Error: Failed to send chunk 20/39 for ServerStorage: Request timeout: /sync/init 2026-08-11T07:01:59.444Z Warning [FLog::CreatorWarning] [SyncManager] init session cleanup failed: Request timeout: /sync/stop 2026-08-11T07:01:59.444Z Info [FLog::CreatorOutput] [SyncWatcher] Stopped ``` Timing from the same log: Studio started 06:59:12Z, the plugin loaded at 06:59:36.563Z, and the init timeout fired **164 s after process start**, with the stop timeout **3.1 s later**. 🔑 **The detail we think matters most: `/sync/stop` times out as well.** The server does not merely fail one request — it stops answering anything, including the cleanup call sent immediately after. That points at a blocked event loop, i.e. something being processed **synchronously**, rather than at a slow or lost individual request. Looking only at `/sync/init` would send you to the wrong place. **3b — Chronology. It is not new, and it is not gone.** All incidents share the form `Failed to send chunk X/Y for <Service>: Request timeout: /sync/init`. | Date | Chunk | Service | Version | Outcome that day | |---|---|---|---|---| | 2026-07-29 | 17/35 | ServerStorage | 2.11.3 | second attempt succeeded | | 2026-07-30 | 20/35 | ServerStorage | 2.11.3 | second attempt succeeded | | 2026-07-31 | 5/8 | **Workspace** | 2.11.3 | server process died; clean full sync after a reboot | | 2026-08-01 | 1/36 and 18/35 | ServerStorage | 2.11.3 | **3 starts / 3 deaths** | | 2026-08-04 | 28/38 | ServerStorage | 2.12.1 | 2 failures / 2 successes | | 2026-08-05 | 18/38 | ServerStorage | 2.12.1 | **0 successes / 2 deaths** | | 2026-08-09 | 6/39 | ServerStorage | 2.12.1 | quoted from the Studio log | | 2026-08-11 | 20/39 | ServerStorage | **2.12.2** | server survived; **sync stayed down** | Two patterns hold across all of them: - **The target is almost always `ServerStorage`** — 8 of 9 timeout incidents. - **The chunk total keeps growing: 35 → 36 → 38 → 39**, while the instance count stayed roughly flat over the same period. **What distinguishes `ServerStorage` here — measured, because our first guess was wrong.** We had assumed it was the service with the most script source. It is not: | Service | Instances | Scripts | Source | |---|---|---|---| | **ServerStorage** | **3803** | 18 | 498.2 KB | | Workspace | 1460 | 0 | 0 KB | | StarterPlayer | 42 | 40 | 375.8 KB | | ServerScriptService | 41 | 40 | **873.6 KB** | | ReplicatedStorage | 21 | 5 | 83.5 KB | `ServerScriptService` carries **1.75× the source bytes** of `ServerStorage` in a comparable script count, and has never produced a timeout. `ServerStorage` holds **~71 % of all instances** in the place (3803 of ~5371) — mostly non-script prop templates. So the failing service is the **instance-heaviest** one, not the source-heaviest. If there is a payload dimension to this, the measurement points at instance/property serialisation rather than at script text — which is also why our own script-byte theory (below) did not survive. **3c — What changed with the 2.12.0 isolation fix, and what did not.** Before it, this condition killed the server process outright (2026-07-31 and 2026-08-01, 3 starts / 3 deaths). Since then the server survives — that fix works. What remains is the failure itself, now on 2.12.2. **3d — The current user-visible damage: the sync stays down silently.** After the 2026-08-11 incident, `manage_sync status_current_place` reported, ~49 minutes later and with no intervening repair: ```json { "state": "idle", "instanceCount": 0, "lastFullSync": null, "activeClientId": null, "fileWatcherActive": false } ``` There was no retry and no recovery; the watcher had stopped and stayed stopped until the sync was started again by hand. Meanwhile the mirror directory still held its 5333 files from the previous session, so **on disk nothing looks wrong** — it is simply frozen. Anything that reads the mirror (static checks, tooling, editors) keeps reporting success against stale content. That silence is the expensive part, more than the failed init itself. --- ## ▼ FIELD 4 — Already ruled out - **"It only affected old versions."** The latest incident is on **2.12.2**, the current release. The pattern spans 2.11.3, 2.12.1 and 2.12.2. - ❌ **Our own script-byte-load theory — we are retracting it here.** We had suspected a folder of script backups in `ServerStorage` (large source payload blocking the event loop). On 2026-08-05 we thinned it: **source 4.84 MB → 0.30 MB (−94 %)** while the instance count moved only 3758 → 3617 (−4 %), and the next sync succeeded on the first attempt where the same day had previously given 0/2. That isolates byte load against instance count — but the timeouts came back on 2026-08-09 and 2026-08-11 with that folder still small (11 files / 465 KB). **So script byte load alone does not explain it,** and we are not asking you to chase it. We mention the numbers only because they show the failure is not simply proportional to payload size. - **Not a place-data problem.** The same place syncs fine at other times — a full sync of 5409 instances / 104 scripts completed the day before the latest incident, into the same directory. - **Not the two issues we filed separately.** The EPERM report concerns `rename` on `observability\changelog\sessions\*.jsonl` with a working retry loop. The archived-place-folder report concerns `ENOENT` in the mirror tree. This one is a request timeout during chunk transfer, with no filesystem error at all. **What we could not determine, and are not guessing about:** what the server is doing during those 164 seconds. That needs a profile or a stack sample inside your process while init is stalling — we can see that it stops answering, not why. If it helps, we can run any instrumentation you suggest the next time it fires; it has recurred often enough that waiting for the next one is realistic. **Suggested direction (yours to judge).** Two things would help independently of the root cause: a timeout on init that leaves the watcher **retrying rather than stopped**, and a surfaced state that distinguishes "mirror is current" from "mirror is frozen because init failed". The second one alone would have saved us the working session we lost to a stale mirror. --- ## ▼ Feedback Several of our earlier reports have shipped — 2.11.2 sync index recovery, 2.11.1 `get_source` line ranges, 2.12.0 Luau write validation plus failure isolation, 2.11.0 `preflight`. The isolation fix is visible in this report's own chronology: the same condition that killed the process in July now leaves it standing.