Description
From the 2026-07-05 performance & architecture review (wasm section, item 4).
Reusable sandboxes now exist as an opt-in (Serve::new().with_max_requests() in Rust). The design doc's rejection of cross-request state ("isolates are not pinned") is no longer absolute. This is a large lever: settings parse, integration registry, signing-key parse, and regex compilation could amortize across requests.
Spike goal
Stand up the opt-in reusable sandbox mode and measure whether it holds up operationally. If it does, the per-request init items (settings pipeline, signing-key parse) and the registry/orchestrator construction amortize across requests, and the OnceLock caches the code already has start earning their keep.
This deserves a measured experiment, not an assumption either way.
Effort / expected effect
- Effort: Spike
- Expected effect: Potentially obsoletes most per-request init cost.
Sequencing item 10 of 10.
Description
From the 2026-07-05 performance & architecture review (wasm section, item 4).
Reusable sandboxes now exist as an opt-in (
Serve::new().with_max_requests()in Rust). The design doc's rejection of cross-request state ("isolates are not pinned") is no longer absolute. This is a large lever: settings parse, integration registry, signing-key parse, and regex compilation could amortize across requests.Spike goal
Stand up the opt-in reusable sandbox mode and measure whether it holds up operationally. If it does, the per-request init items (settings pipeline, signing-key parse) and the registry/orchestrator construction amortize across requests, and the
OnceLockcaches the code already has start earning their keep.This deserves a measured experiment, not an assumption either way.
Effort / expected effect
Sequencing item 10 of 10.