Repository navigation
docs(code-index): design a smaller decoded generation - #4560
Conversation
The sealed codec already writes per-chunk revisions, the projection request, and clone token streams compactly. The decoded generation expands them, so the 3.24 GB that issue #4506 measured is a decoded representation cost, and no sealed-format revision is needed. The design orders four decoded-form changes, keeps format revision 21, explains why a later byte change rebuilds from source instead of migrating, and records how to measure each change. It also records the current cold-pass peaks: 1.58 GB on this repository and 2.45 GB on t3code, with the peak now in the uncharged interactive catalog warm. Refs #4506 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
There was a problem hiding this comment.
🔍 Devin Review: 1 flag
Not posted on this PR by your GitHub settings — view it in Devin Review. (Configure)
|
Corrected the documentation findings in 3b051a5. The maintained measurement procedure uses the supplied checkout and an isolated temporary profile, not a project-data copy. Decode admission now distinguishes retained bytes from measured peak demand and explicitly leaves the first-unmeasured-decode gap open. Document roster order follows the existing file rows, and the incorrect serde feature statement is removed. An independent final source review passed the exact committed document blob 8337279c8ca6abd6fa0057a3e934618463daf63a. This is documentation-only; no application build or new benchmark claim is made. New CI remains pending. |
Refs #4506
Problem
Issue #4506 reported a cold build that held a 3.24 GB decoded generation that admission did not charge. That build peaked at a 4.11 GB footprint on this repository.
What master already fixes
#4510 removed that hold. Cold construction now writes segments to the store preparation and charges what it retains as
code_index.construction.v1. A refused build parks until its demand fits.publish_atomicallyandCodeIndexSealedPublicationV1::encodeno longer exist. Those two functions are the frames the live beta.92 daemon is stuck in.I measured the release daemon at
bbfcad7c03in an isolated profile with a fresh clone. Each journey ran once.During construction, the reservations stayed above occupancy. The remaining peak comes after publication, from the interactive catalog warm, which takes no reservation. That gap is now #4559.
This PR
This PR adds a design doc and changes no code. The doc is
docs/plans/CODE-GENERATION-RESIDENT-SIZE-DESIGN.md, and it covers the sealed-codec candidates from the issue. The sealed codec already writes per-chunk revisions, projection rows, and clone token streams compactly. The 3.24 GB comes from expanding them during decode. The doc orders four decoded-form changes and keeps format revision 21. It explains why a later byte change rebuilds from source instead of migrating, and it gives the measurement method.🤖 Generated with Claude Code