Problem
Filesystem access under memory_dir assumes local disk.
Artifact reads and writes use tokio::fs, but the checks around them don't. Seven .exists() and two .canonicalize() calls in persistence.rs run synchronously inside async fns, blocking a runtime thread instead of the blocking pool. load_session_manifests stats the session dir, lists it, then stats each entry twice, so at the default 20 runs per session that's ~40 synchronous fs calls per turn.
The scratchpad writes intercepted tool output with a blocking std::fs::write on the request thread (scratchpad/wrapper.rs). The comment there says to revisit it "if profiling shows the sync write on the hot path".
persistence_drain_timeout_ms defaults to 2000, sized for sub-millisecond writes. On timeout the run continues with a warning, so a manifest can reference tool output artifacts that haven't landed.
Context
Artifacts hold prior worker evidence and prior tool calls. A follow-up turn needs them, and with no session affinity on a multi-pod deploy they're just not there. Sharing them across pods means putting memory_dir on network storage, where none of these calls are free.
#226 closed on local-disk measurements: 0ms drain across nine sampled iterations. It left the case open for large-artifact workloads that show drain time in a profile.
Notes
Config and single-pod behavior unchanged. #218's instrumentation would show the drain cost.
Problem
Filesystem access under
memory_dirassumes local disk.Artifact reads and writes use
tokio::fs, but the checks around them don't. Seven.exists()and two.canonicalize()calls inpersistence.rsrun synchronously inside async fns, blocking a runtime thread instead of the blocking pool.load_session_manifestsstats the session dir, lists it, then stats each entry twice, so at the default 20 runs per session that's ~40 synchronous fs calls per turn.The scratchpad writes intercepted tool output with a blocking
std::fs::writeon the request thread (scratchpad/wrapper.rs). The comment there says to revisit it "if profiling shows the sync write on the hot path".persistence_drain_timeout_msdefaults to 2000, sized for sub-millisecond writes. On timeout the run continues with a warning, so a manifest can reference tool output artifacts that haven't landed.Context
Artifacts hold prior worker evidence and prior tool calls. A follow-up turn needs them, and with no session affinity on a multi-pod deploy they're just not there. Sharing them across pods means putting
memory_diron network storage, where none of these calls are free.#226 closed on local-disk measurements: 0ms drain across nine sampled iterations. It left the case open for large-artifact workloads that show drain time in a profile.
Notes
Config and single-pod behavior unchanged. #218's instrumentation would show the drain cost.