fix(host-mcp): evict idle MCP sessions instead of leaking them - #1685
Merged
RhysSullivan merged 7 commits intoAug 28, 2026
Merged
Conversation
The in-process session store keyed transports, servers, owners, and engines by session id and only ever deleted an entry on `onsessionclosed`, which the SDK fires on `DELETE /mcp`. Nothing sends that DELETE: the client SDK's `transport.close()` aborts locally and puts nothing on the wire, a crashed client cannot send it, and `enableJsonResponse` leaves no stream whose teardown could stand in for it. Every initialize therefore pinned an McpServer, its tool registry, and an ExecutionEngine until the process exited. Measured against ghcr.io/usefulsoftwareco/executor-selfhost:1.5.42, 500 sessions opened without a DELETE grow RSS by 346 MiB (709 KiB each, linear, no plateau); the same 500 with a DELETE grow it by 13 MiB. Stamp each session on create and on every forwarded request, then sweep on a timer and dispose anything idle past the TTL. Eviction is what the streamable HTTP spec allows a server to do, and the store already renders an unknown id as the existing "not-found" (404 -32001), which is a client's cue to re-initialize.
The store's idle window is only useful if an operator can tune it: a client that cannot tolerate re-initializing needs a longer TTL, and diagnosing one needs eviction off entirely (0). Parse it the same way as EXECUTOR_SANDBOX_TIMEOUT_MS, refusing to boot on a malformed value.
Resolutions: - in-memory-session-store.ts: main dropped requestStateSigningKey; keep that removal alongside the branch lastSeen/touch/sweep additions. - in-memory-session-store.test.ts: main rewrote this file down to a single case, deleting the buildMcpServer and eliciting-engine helpers the branch test used. Rebuild the eviction case on createExecutorMcpServer and drive sweepIdleSessions with an explicit instant instead of real sleeps.
…ationale makeSelfHostMcpSeams called loadConfig() inside the factory, hiding an env read behind construction and moving a boot-time throw (loadConfig refuses a malformed EXECUTOR_MCP_SESSION_IDLE_TTL_MS) off the boot path. The app already has the resolved config, so pass it in. The store comment claimed enableJsonResponse means no long-lived stream exists. It only governs how a POST carrying requests answers; the bare 202 for notifications/initialized still cues the client to open the GET stream, so nearly every session holds one. Eviction ignores it on purpose, which is what cloud already does past its running-lease ceiling.
… down The idle sweep read only a last-seen stamp, written before the store awaits transport.handleRequest. A call slower than the idle window was therefore indistinguishable from an abandoned session, and the sweep closed the transport, the server, and the engine underneath the request still using them. Count the requests inside handleRequest per session, skip a session with any, and restamp when a call ends so idleness measures from completion. Disposal deleted the engine reference without running engine.shutdown, so the detached sandbox fibers a paused execution holds kept running - and kept querying the host database handle - after the session was gone. Every disposal path now shuts the engine down: sweep eviction, the dispose seam, store close, onsessionclosed, and the eager close of a transport that never minted an id. A close failure was swallowed whole, which made a leaked handle invisible. Keep it best-effort, but log it at warning with the session id and the handle. Wire the store close hook into the self-host server. startServer discarded closeDb, so a graceful shutdown left every live session and the libSQL handle open; only the test web-handler path ever released them.
Closed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes #1684.
What this does
The in-process MCP session store now evicts sessions that have gone idle, instead of holding every session it ever created until the process exits.
Each session gets a last-seen stamp on create and on every forwarded request. A timer sweeps the stamps and disposes anything past
sessionIdleTtlMs, which defaults to 30 minutes. The self-host exposesEXECUTOR_MCP_SESSION_IDLE_TTL_MSso an operator can widen the window, or set0to turn eviction off.Why
onsessionclosedfires onDELETE /mcp, and nothing sends that DELETE.StreamableHTTPClientTransport.close()clears a timer, aborts a localAbortController, and callsonclose. It puts nothing on the wire. OnlyterminateSession()sends the DELETE, andClient.close()does not call it.SIGKILLed client cannot send it at all.enableJsonResponse: truemeans there is no long-lived stream whose teardown could stand in for it.So a session that the client has finished with stays in
transports,servers,owners, andengines, pinning anMcpServer, its tool registry, and anExecutionEngine. Across 12 hours of production logs on my self-host I counted zeroDELETE /mcprequests against 157 new sessions an hour.Evicting is what the streamable HTTP spec allows a server to do, and the store already behaves correctly when it happens: an unknown session id falls through to the existing
"not-found", which the envelope renders as404 -32001, and the client re-initializes.What I ran
Both scripts are in the gist on #1684. The load is identical in every arm: N sessions, client never sends
DELETE.Retention. Fill the heap, idle long enough for the sweep to drain, then repeat the identical load. RSS in MiB:
1.5.421.5.42keeps climbing across batches and never gives the memory back. This branch returns below its own starting point and reaches the same ceiling on the second batch rather than a higher one. The in-batch rise is the live working set: 400 sessions opened in about two minutes against a 15-second TTL are legitimately alive at once.Eviction behaves as designed. With a 20-second TTL, a session answers
200immediately and404 -32001after 75 seconds.Gates.
in-memory-session-store.test.ts3/3 with a new test covering evict-the-idle-one, keep-the-busy-one, and 404-after-eviction.apps/host-selfhostmcp and config suites 17/17.tsgo --noEmitandoxlintclean on every changed file.packages/hosts/mcpis 202/203. The one failure isstdio-integration.test.ts, which fails the same way on unpatchedmainataff1f394, so it is not from this change.What I did not check
I have not run the full monorepo suite, only the two packages this touches. I have not exercised the Cloudflare Durable Object store, which has its own lifecycle and is untouched here.
Choices worth a look
30 minutes as the default. Long enough that a polling client (mine reconnects about every five minutes) never trips it, short enough that an abandoned session does not outlive the working day. Happy to change it.
The sweep swallows its own failures. It follows
ignoreClose: a failed sweep must not surface as an unhandled rejection in a host that is otherwise healthy. The timer isunref'd so it never keeps a process alive by itself.Eviction closes the transport as well as the server, unlike the
onsessionclosedpath which passes only{ server: true }. An evicted session's transport has no other owner, and leaving it open would hold the handles the eviction exists to release. I left the DELETE path alone rather than widen the scope here.