Repository navigation
[bug] query_graph hangs on whole-graph OPTIONAL MATCH (no execution timeout / degradation) on large graphs #601
Description
Version: codebase-memory-mcp v0.8.1
Project: real-world repo, 93,315 nodes / 488,306 edges
Summary
A query_graph Cypher that uses an unbounded OPTIONAL MATCH over the full node set does not return — it appears to hang rather than hit the documented 100k-row ceiling, because the optional expansion blows up before any rows are produced. There is no execution-time guard, so the caller just waits.
Repro
On a 90k+ node graph:
MATCH (a)
OPTIONAL MATCH (a)-[:CALLS]->(b)
RETURN a.qualified_name, count(b)→ does not return in a reasonable time (no partial result, no error).
By contrast, a bounded/directed rewrite returns instantly:
MATCH (f:Function)-[:CALLS]->(g)
WHERE f.transitive_loop_depth >= 3
RETURN f.qualified_name, count(g)
ORDER BY count(g) DESC
LIMIT 20Expected
query_graphshould enforce an execution timeout (wall-clock), not only a row ceiling, and return a clear error (or partial result) instead of hanging.- Ideally: a planner guard that warns/aborts on unbounded
OPTIONAL MATCHagainst the full graph, suggesting a directed/LIMITed rewrite. - Document the row ceiling vs time limit distinction in the tool description.
Workaround (consumer side)
Avoid whole-graph OPTIONAL MATCH: pre-filter the driving set with WHERE, use directed MATCH + count() for optional-cardinality, always LIMIT, and chunk by file_pattern for large sweeps. But a server-side timeout would prevent the foot-gun entirely.
Activity
DeusData commentedon Jun 29, 2026
Thanks. I’ve triaged this as Cypher stability/performance under #593 and #594. The likely fix needs execution guardrails, such as timeout/budgeting or bounded degradation for optional expansions before row materialization.
rodmank commentedon Jun 30, 2026
Independently corroborating on Linux / WSL2 (x86-64), v0.8.1 — same no-timeout/no-degradation behavior, and notably it triggers even on a small graph, not just 90k+ nodes.
Project: 5,996 nodes / 14,201 edges, via
codebase-memory-mcp cli query_graph(fresh process per call):→ completes but takes 8.5 s wall / 493 MB RSS (vs sub-ms for bounded depth-3 paths). Cost grows unboundedly well before the 100k-row ceiling, consistent with the missing execution-time guard described here.
Unsupported-syntax constructs (
EXISTS {…},RETURN ppath variables) return clean errors immediately — so the foot-gun is specifically unboundedOPTIONAL MATCH/ variable-length-path / combinatorial joins, not parsing.Client-side impact: when such a query exceeds the memory budget the server process dies, which over the MCP stdio transport surfaces as
MCP error -32000: Connection closed. The client (Claude Code) auto-reconnects (~12 ms) and retries the same query, which crashes again — a reconnect storm that to the end user looks exactly like "the server hangs / doesn't answer." A server-side wall-clock timeout + clean error (as proposed here) would resolve both the hang and the reconnect storm.+1 for the execution-time guard and the planner warning on unbounded whole-graph expansion.
rudi193-cmd commentedon Jul 5, 2026
Consumer-side mitigation pattern (Willow fleet audit)
We hit the crash/hang class during a June structural audit and shipped a bounded wrapper (
sap/cbm_facade.pyin willow-2.0) that:count(*),coalesce(),<-, etc.) before it reaches the serverLIMITwhen missing onquery_graphWILLOW_CBM_TIMEOUT_S, default 30s) so a wedged query does not kill the IDE stdio sessionDead-code style queries should use forward-only
MATCH (caller)-[:CALLS]->(f)with explicitLIMIT, not whole-graphOPTIONAL MATCHanti-joins — matches the bounded query in #627.Not a substitute for server-side timeout/degradation, but documents what downstream integrators are doing until #601/#627 are fixed. Related: #873 (DISTINCT ordering still broken on v0.8.1 despite #237).
Merge pull request #1109 from DeusData/fix/601-cypher-exec-deadline
fix(cypher): add a wall-clock execution deadline to abort runaway que…
fix(cypher): add a wall-clock execution deadline to abort runaway que…
fix(cypher): expansion materializes every matched row — the cap falsi…