Repository navigation
perf: bind generic interaction append payloads - #10122
Conversation
|
Archestra reviewed Archestra code reviewNothing blocks the merge · reviewed Address #1 first: run Since Since f386794, the only source change is removal of the one-off interaction-append workflow; The SQL approach remains sound: Focused verification passed: backend type checking, migration journal and lint checks, Drizzle metadata validation, 11 migration-upgrade tests, seven existing append-model tests, and the disposable PostgreSQL 18 test. That test reported medians of 2.50→2.06 ms (~64 KiB) and 27.26→24.53 ms (~1 MiB), consistent with the optimization’s direction; these are local client-observed timings, not a production latency guarantee.
Review notesReviewed
Observations
Tests
The focused tests cover the changed generic path; no changed behavior lacked a runtime check. The new PostgreSQL script does not repeat all four credential-shape combinations covered by the existing PGlite suite. Not covered
History: 6 jobs
|
There was a problem hiding this comment.
Changes requested, resolved in f386794 · Archestra review of dd7dd41 · 1 finding: 1 medium. Summary
The review of f386794 finds nothing that blocks the merge.
There was a problem hiding this comment.
Nothing blocks the merge · Archestra review of 409f9b1 · 1 finding: 1 low. Summary
| const current = definitions.at(-1); | ||
| const profileId = "00000000-0000-4000-8000-000000000001"; | ||
|
|
||
| test("append upgrade preserves encoded values, defaults, replay and accounting", async (t) => { |
There was a problem hiding this comment.
1. [low] interaction-append-postgres.test.mjs has no CI invocation
The PostgreSQL regression test added for the bound generic path is no longer referenced by any workflow or package test script.
The latest revision removes the dedicated interaction-append workflow, which was the only caller of this test. The current workflows and package scripts contain no invocation of interaction-append-postgres.test.mjs; the existing model suite uses PGlite instead.
If a later migration changes generic native casting in a way supported by PGlite but rejected by PostgreSQL, the retained CI checks can pass without executing this regression test. The actual database behavior this PR relies on is therefore unprotected despite keeping the test file.
Recommendation:
Invoke this test from an existing migration-validation workflow with an isolated PostgreSQL service and its required connection variable.
🤖 I have created a release *beep* *boop* --- ## [1.5.0-rc.21](archestra-ai/archestra@platform-v1.5.0-rc.20...platform-v1.5.0-rc.21) (2026-11-04) ### Features * **agent-runtime:** audit live terminal attaches and confirm Ctrl+C before it reaches the agent ([archestra-ai#10134](archestra-ai#10134)) ([536445b](archestra-ai@536445b)) * **guardrails:** Break the glass settings let blocked calls through with a yell, unattended runs included ([archestra-ai#10126](archestra-ai#10126)) ([dd9fa12](archestra-ai@dd9fa12)) * **guardrails:** connect ChatGPT desktop with native plugin setup ([archestra-ai#10051](archestra-ai#10051)) ([bc24cbd](archestra-ai@bc24cbd)) * **guardrails:** rename Soft block to Other blocked calls; Yell tool first ([archestra-ai#10174](archestra-ai#10174)) ([0f46b43](archestra-ai@0f46b43)) * **llm-proxy:** rename Model Router to Unified Model API ([archestra-ai#10138](archestra-ai#10138)) ([25cc30c](archestra-ai@25cc30c)) * **llm:** connect to a model pool with the LLM Proxy connect steps ([archestra-ai#10157](archestra-ai#10157)) ([11c7f9d](archestra-ai@11c7f9d)) * **model-pools:** support Gemini and Vertex destinations ([archestra-ai#10088](archestra-ai#10088)) ([dc1ab2d](archestra-ai@dc1ab2d)) * **models:** clean up unconfigured request-discovered entries ([archestra-ai#10150](archestra-ai#10150)) ([981f2a4](archestra-ai@981f2a4)) * **sre:** migrate agent to built-in object permissions ([archestra-ai#10146](archestra-ai#10146)) ([699b428](archestra-ai@699b428)) ### Bug Fixes * **agents:** clarify agent creation choices and responsive layout ([archestra-ai#10172](archestra-ai#10172)) ([9b6af74](archestra-ai@9b6af74)) * **agents:** return chat-capable built-in agents in search ([archestra-ai#10136](archestra-ai#10136)) ([d3dda50](archestra-ai@d3dda50)) * **backend:** scope user agent runtime gateway tokens and enforce start_agent_run subagent allowlists ([archestra-ai#10119](archestra-ai#10119)) ([f4d1ba7](archestra-ai@f4d1ba7)) * **backend:** tests install the published native addons when a checkout has none ([archestra-ai#10135](archestra-ai#10135)) ([bff6802](archestra-ai@bff6802)) * **bedrock:** guard oversized Converse prompts before inference ([archestra-ai#10143](archestra-ai#10143)) ([ac769a3](archestra-ai@ac769a3)) * **connect:** count Desktop instruction code points ([archestra-ai#10121](archestra-ai#10121)) ([a8bf336](archestra-ai@a8bf336)) * **frontend:** explain why personal Claude subscription requires archestra-claude-code command ([archestra-ai#10179](archestra-ai#10179)) ([ed95113](archestra-ai@ed95113)) * **guardrails:** keep human review by default after the Break the glass upgrade ([archestra-ai#10164](archestra-ai#10164)) ([e8e9ffc](archestra-ai@e8e9ffc)) * **knowledge:** configure environments for uploaded files ([archestra-ai#10145](archestra-ai#10145)) ([e418a87](archestra-ai@e418a87)) * **llm-proxy:** abort provider requests when clients disconnect ([archestra-ai#10092](archestra-ai#10092)) ([7adb963](archestra-ai@7adb963)) * **llm-proxy:** block unpriced vLLM requests under spend limits ([archestra-ai#10151](archestra-ai#10151)) ([95f533e](archestra-ai@95f533e)) * **llm-proxy:** show unknown usage when streams end early ([archestra-ai#10177](archestra-ai#10177)) ([383ad0c](archestra-ai@383ad0c)) * **mcp:** expose gateway access flags and agent skill policies ([archestra-ai#10116](archestra-ai#10116)) ([14afd2b](archestra-ai@14afd2b)) * **mcp:** unwrap run_tool wrappers around nullable params ([archestra-ai#10155](archestra-ai#10155)) ([1d72de4](archestra-ai@1d72de4)) * **models:** block new requests to hidden models ([archestra-ai#10087](archestra-ai#10087)) ([b9fa196](archestra-ai@b9fa196)) * **observability:** log LLM proxy pre-handler denials ([archestra-ai#10178](archestra-ai#10178)) ([ca7e990](archestra-ai@ca7e990)) * **openappa:** let Claude Code read its own background command output ([archestra-ai#10139](archestra-ai#10139)) ([e1e1980](archestra-ai@e1e1980)) * **openappa:** support GitHub Enterprise in the github battery and repin OpenAPPA ([archestra-ai#10133](archestra-ai#10133)) ([1fcac6c](archestra-ai@1fcac6c)) * **teams:** honor effective Enterprise entitlement for Team Sync ([archestra-ai#10153](archestra-ai#10153)) ([ab73d4b](archestra-ai@ab73d4b)) * **teams:** require update permission for MCP team management ([archestra-ai#10165](archestra-ai#10165)) ([6abae54](archestra-ai@6abae54)) * **teams:** roll up budgets and delegate member key management ([archestra-ai#10175](archestra-ai#10175)) ([548a726](archestra-ai@548a726)) ### Performance Improvements * bind generic interaction append payloads ([archestra-ai#10122](archestra-ai#10122)) ([265414b](archestra-ai@265414b)) ### Code Refactoring * **dev:** serialize checks and restore portable compiler caches ([archestra-ai#10104](archestra-ai#10104)) ([2a1bffd](archestra-ai@2a1bffd)) ### Miscellaneous Chores * **openappa:** pin OpenAPPA with the annotator label-guide fixes ([archestra-ai#10131](archestra-ai#10131)) ([fde9743](archestra-ai@fde9743)) * update agent runtime CLIs ([archestra-ai#10125](archestra-ai#10125)) ([0e71d0b](archestra-ai@0e71d0b)) --- This PR was generated with [Release Please](https://github.com/googleapis/release-please). See [documentation](https://github.com/googleapis/release-please#release-please). --------- Co-authored-by: archestra-ci[bot] <222894074+archestra-ci[bot]@users.noreply.github.com>
Generic interaction appends currently embed encoded values into generated INSERT SQL. Large request/response bodies therefore enlarge the SQL PostgreSQL must parse on every call. Bind the payload and generate only validated column references and native casts, preserving omitted-column defaults, cached common shapes, retry semantics and activity accounting.
The migration only replaces a function; it adds no indexes and rewrites no interaction rows. The generated snapshot records the unchanged schema.
Validation: the PostgreSQL 17 CI run passed cross-release migration-history compatibility, migration metadata validation, and the function-upgrade regression checks: encoded JSON/SQL nulls, precision, defaults, existing-row replay, invalid payload rejection and exact activity counts. The migration-check regression verifies a valid stable worktree fails in main mode and passes in release mode. The dedicated validation workflow used for these measurements has been removed from this PR; existing CI retains the migration-check regression. No local test suites were run; the broader merge-queue suite has not run.
The initial synthetic generic-append benchmark improved median timing from 2.59 to 2.22 ms (~64 KB) and from 26.77 to 22.82 ms (~1 MB). These are isolated generic-path measurements, not a claim to explain all append latency.
The provisional release/1.4 backport #10123 was closed before merge. Introducing its migration after main-only migrations would violate the migration prefix rule and cause later upgrades to skip required migrations; do not backport this migration independently.