Project
Release process
The checks every release passes before it is published.
What has to pass before a version is tagged, and how each check is run. CI (.github/workflows/ci.yml) covers checks 1, 2 and 9 on every push, on x86_64 and arm64. The rest are run by hand on the release candidate's exact commit, and their results go in docs/PROGRESS_LOG.md.
| # | Check | How | Pass |
|---|---|---|---|
| 1 | Lints | cargo clippy --workspace --all-targets -- -D warnings | no warnings |
| 2 | Test suite | cargo test --release --workspace --no-fail-fast, on x86_64 and arm64 | every test passes |
| 3 | Process chaos | cargo test --release -p storage-server --test process_chaos, 40 times, on x86_64 and arm64 | 40 of 40 on each |
| 4 | Soak, single group | STORINE_SOAK_SECS=7200 STORINE_SOAK_WINDOW_SECS=600 cargo test --release -p storage-server --test process_chaos soak_single_group -- --ignored --nocapture, 8 vCPU and 16 GB | the test passes: balances conserved, every key linearizable, memory and descriptors under their ceilings and not climbing |
| 5 | Soak, range cluster | as 4, with soak_range_cluster | as 4 |
| 6 | Undefined behaviour in unsafe code | Miri: MIRIFLAGS="-Zmiri-strict-provenance -Zmiri-disable-isolation" cargo +nightly miri test -p storage-codec --lib -- scratch checksum. AddressSanitizer: ASAN_SYMBOLIZER_PATH=$(command -v llvm-symbolizer) LSAN_OPTIONS=suppressions=$PWD/tools/lsan.supp:fast_unwind_on_malloc=0 RUSTFLAGS=-Zsanitizer=address cargo +nightly test -Zbuild-std --target <host> --no-fail-fast -p storage-codec -p storage-mmap -p storage-wal --lib -- --skip freed_memory_can_be_released | no report. The skipped test checks that glibc's malloc_trim returns memory, which it cannot under AddressSanitizer's allocator; the suppression covers a test that leaks a segment on purpose (it simulates a crash). The suppressions path must be absolute (some storage-mmap tests change the working directory), a symbolizer must be installed (llvm; suppressions match function names), and the full unwinder must be on: the fast one cut a stack of that test short and its leak was reported. Install nightly with rustup toolchain install nightly -c miri,rust-src. |
| 7 | Benchmarks | cargo run --release -p storage-benches --bin bench-runner, and benches/kv-compare against LMDB and RocksDB (docs/BENCHMARK_REPORT.md section 8) | no unexplained regression against the last release |
| 8 | SDKs | each SDK's suite (commands in tools/modal/sweep.py, SDK_SUITES) against the release storage-server (STORINE_SERVER): conformance vectors, end to end on one node and a 3-node cluster, and TLS (storines://, certificates from gen-ca / gen-cert) | every suite passes |
| 9 | Binary size | the release build for the baseline CPU on Linux x86_64 (tools/release/build-dist.sh; CI builds the same on every push) | under the RULE-MEM-03 ceilings: 5.5, 1.75, 8 and 1 MiB (storine-console) |
| 10 | Release notes | the version's entry in CHANGELOG.md: what changed, what it fixes, known limitations, and which formats and protocols changed (docs/COMPATIBILITY.md) | written, with numbers from checks 7, 4–5 and 9 |
unsafe is allowed only in storage-codec and storage-mmap (docs/RULES.md), which is why check 6 covers those crates. It also covers storage-wal, whose tests drive the codec's frame pool from many threads.
Costs on AWS on-demand (us-east-1, 2026-11): checks 2 and 3 on a c7i.xlarge (x86_64) or c7g.xlarge (arm64) take about 1.5 h, about $0.30. Each soak on a c7i.2xlarge takes about 2.5 h with the build, about $0.90. Launch instances with --instance-initiated-shutdown-behavior terminate and a shutdown -h +N backstop, and terminate them when the job reports.