Status: architecture first
Following the maintainer's recommendation in #2678, address the architectural installation cost first without adding a decompression backend. #2676 is the owning implementation PR for single-pass streaming and trusted build-generated per-file hashes/metadata that allow valid warm verification without decompressing the archive. Retain the existing miniz_oxide backend and preserve content integrity, permissions, repair behavior, bounded memory, and concurrency guarantees.
The backend experiment in #2678 is now a draft and is deferred. Reconsider zlib-rs only after the architectural work lands and fresh matched cold/warm/repair benchmarks plus release binary-size measurements demonstrate a worthwhile remaining benefit. Earlier measurements against the eager installer are historical evidence, not post-architecture results.
Problem
At upstream revision 9553d5224c73df2d02aee5ec01eeb8353acc736a, the Rust bundled runtime installer traverses the same embedded gzip archive repeatedly and reconstructs large entries even during warm verification. Installer-only measurements show avoidable decompression and allocation costs.
Reproduction
Build the Rust SDK in release mode with its default bundled-cli feature and the pinned shipped runtime. In a fresh process with an isolated HOME, time the public github_copilot_sdk::install_bundled_runtime() API. Repeat with a fresh process against the same valid installation. Do not launch a CLI process, authenticate, or make a model request. Compare cold installation, valid warm verification, same-size corrupted runtime.node, and runtime.node truncated to 1024 bytes.
Evidence required
Use matched source/compiler/profile/runtime/archive identities and repeated alternating before/after processes. Compare output digests, sizes and modes; report latency, CPU, peak and retained memory, and release binary size. Prove valid warm verification reads no archive bytes while still checking installed content against trusted build-generated metadata rather than mutable cache markers.
Deferred backend tradeoff
flate2 documents the byte-slice adapter and backend options at https://docs.rs/flate2/1.1.9/flate2/. zlib-rs requires no C compiler but adds unsafe dependency internals. Any future backend proposal must justify that tradeoff with evidence on the completed architecture, not the old repeated-traversal implementation.
Status: architecture first
Following the maintainer's recommendation in #2678, address the architectural installation cost first without adding a decompression backend. #2676 is the owning implementation PR for single-pass streaming and trusted build-generated per-file hashes/metadata that allow valid warm verification without decompressing the archive. Retain the existing miniz_oxide backend and preserve content integrity, permissions, repair behavior, bounded memory, and concurrency guarantees.
The backend experiment in #2678 is now a draft and is deferred. Reconsider zlib-rs only after the architectural work lands and fresh matched cold/warm/repair benchmarks plus release binary-size measurements demonstrate a worthwhile remaining benefit. Earlier measurements against the eager installer are historical evidence, not post-architecture results.
Problem
At upstream revision
9553d5224c73df2d02aee5ec01eeb8353acc736a, the Rust bundled runtime installer traverses the same embedded gzip archive repeatedly and reconstructs large entries even during warm verification. Installer-only measurements show avoidable decompression and allocation costs.Reproduction
Build the Rust SDK in release mode with its default
bundled-clifeature and the pinned shipped runtime. In a fresh process with an isolated HOME, time the publicgithub_copilot_sdk::install_bundled_runtime()API. Repeat with a fresh process against the same valid installation. Do not launch a CLI process, authenticate, or make a model request. Compare cold installation, valid warm verification, same-size corrupted runtime.node, and runtime.node truncated to 1024 bytes.Evidence required
Use matched source/compiler/profile/runtime/archive identities and repeated alternating before/after processes. Compare output digests, sizes and modes; report latency, CPU, peak and retained memory, and release binary size. Prove valid warm verification reads no archive bytes while still checking installed content against trusted build-generated metadata rather than mutable cache markers.
Deferred backend tradeoff
flate2 documents the byte-slice adapter and backend options at https://docs.rs/flate2/1.1.9/flate2/. zlib-rs requires no C compiler but adds unsafe dependency internals. Any future backend proposal must justify that tradeoff with evidence on the completed architecture, not the old repeated-traversal implementation.