feat(solana-indexer) PR 14: binary wiring

[squadgazzz]

Description

The binary comes alive: a TOML configuration (same shape as the solana-driver's, reusing configs::shared), the database pool, the RPC client, the ingester-to-decoder channel, and a shutdown signal, all assembled in run::start. The CLI takes --config plus a --start-slot override for the first subscription (bounded by the provider's replay window). The process runs until the stream loop or the decoder stops or a signal arrives.

The stream loop owns the restart policy. A dropped stream reconnects and resumes one past the last indexed slot. When that resume is rejected, which usually means the slot fell out of the provider's replay window (~150 slots), the loop logs an error and resubscribes from the live tip instead of crash-looping. The gap stays unindexed until the MVP backfill (BE-204).

SolFlow has no program yet, so its subscription filter only exists when --solflow-program is set, and the decoder sees the settlement program in its place, which makes the SolFlow branch unreachable.

Consuming the modules also fired the planted dead-code expectations (Decoder, Postgres::new, Rpc::new), so those attributes are gone.

Two named gaps, both out of scope here: the latest_chain_slot counter has no reader and solana.chain_tip has no writer yet (that pairing lands with the autopilot's deadline work), and there is no liveness endpoint until the metrics PR (BE-73).

Changes

• config: TOML file loading with an example.toml kept parseable by a test
• run::start: config, observability init, pool, RPC client, channel, spawned decoder, stream loop, shutdown via tokio::select!
• Ingester::serve takes a Resume (watermark or live tip) and an optional SolFlow program
• Deleted the temporary Send-assertion helper, the real spawn replaces it

How to test

Existing tests. The real test is running the binary against on-chain data, which is the next step after this stack lands.

Related issues

Resolves BE-75.

[linear-code]

BE-75

[squadgazzz]

@claude please review

[claude]

Claude finished @squadgazzz's task in 4m 58s —— View job

---

PR Review: binary wiring

Reviewed the full diff against origin/solana-indexer/PR13-onchain-orders. The wiring is clean and closely mirrors the established solana-driver pattern (TOML config via configs::shared, observe init, tokio::select! shutdown). Config handling, the Option<Pubkey> refactor for the not-yet-existing SolFlow program, the conditional subscription filter, and the removal of the dead-code placeholders all look correct and well-documented. The Resume state machine (Watermark → LiveTip fallback on rejected resume) reads sensibly and self-heals as the watermark advances.

One issue worth addressing, posted inline:

• run.rs (tokio::select!) — a decoder PersistenceError can be silently swallowed. decoder.run() returns a Result, so a DB failure makes it exit with Err without panicking; that causes serve to return Ok(()) and stream_loop to break at the same moment decoder_task is ready. Since select! polls arms in random order, the root-cause error is only logged about half the time — the rest of the time it logs the generic "stream loop stopped". Suggest surfacing the decoder result deterministically after the loop rather than via the race.

Nothing else blocking — the remaining behavior (live-tip gaps until backfill BE-204, no graceful decoder drain on shutdown) is already acknowledged in the PR description and acceptable for the MVP.
· solana-indexer/PR14-wiring

[claude]

Claude finished @squadgazzz's task in 3m 7s —— View job

---

PR Review: binary wiring

• Read the PR diff against origin/solana-indexer/PR13-onchain-orders
• Analyze the changed files
• Review run::start wiring, Resume state machine, config loading, and the Option<Pubkey> refactor
• Verify the previously-flagged decoder-exit race

---

Re-reviewed the full diff. Since my last review, the decoder-exit concern I raised has been fixed (commit e4ef70bf4): the tokio::select! now awaits and logs decoder_task deterministically after stream_loop breaks, instead of relying on the random arm-polling order. I traced the terminal states to confirm the fix is sound:

• Ingester::serve returns Ok(()) only when the decoder dropped its rx (run.rs:103 → ingester.rs:114-134). By then decoder.run() has returned and the spawned task is complete, so (&mut decoder_task).await at run.rs:127 resolves promptly rather than hanging — and the shutdown arm can't be starved.
• A returned PersistenceError now surfaces via arm 2 (result = &mut decoder_task), which fires directly on decoder death even while the ingester is idle in stream.next(). 👍

Everything else holds up:

• Resume state machine self-heals correctly. Watermark→LiveTip on a rejected resume (no crash-loop), and after a LiveTip stint the watermark has advanced back inside the replay window, so a later reconnect resumes cleanly from Watermark.
• Config (config.rs) mirrors the solana-driver pattern; deny_unknown_fields, the optional-pubkey deserializer paired with #[serde(default)], and the example.toml parseability test are all correct and validated by load_example_toml.
• Option<Pubkey> SolFlow refactor is threaded consistently through the decoder, relevant_instructions, the subscription filter (only inserted when set), and the updated tests.
• Dead-code placeholder removals (Decoder, Postgres::new, Rpc::new) and the deleted assert_serve_future_is_send helper are all justified now that the real spawn exists.

No blocking issues. The acknowledged gaps (live-tip gap until BE-204 backfill, no reader for latest_chain_slot / no solana.chain_tip writer until the autopilot deadline work, no liveness endpoint until BE-73, no graceful decoder drain on shutdown) are all noted in the PR description and fine for the MVP. LGTM.
· solana-indexer/PR14-wiring