yellowstone-reconcile

Reconcile complete Solana blocks across independent Yellowstone connections.

The problem

Solana's Alpenglow integration guidance points consumers at per-connection block reconstruction, then stops short of the next layer:

Run one buffer per connection. bank_id means nothing outside the stream it arrived on. [...] Reconciling across connections is still yours to write.

solana.com/upgrades/alpenglow

The same page warns that merging on (slot, bank_id) across streams fails silently and hands back blocks fused from unrelated banks. Alpenglow activated on mainnet on 21 October 2026, and the Foundation marks it a breaking change that requires indexing changes. This library is that missing layer.

What it does

• Takes one upstream BlockStream per connection and never mixes their events.
• Reconciles sealed blocks on blockhash, parent and counts, not on bank_id.
• Requires a configurable number of distinct providers before emitting.
• Quarantines disagreement and reports conflicts explicitly, never selecting silently.
• Keeps provider identity stable across reconnects, so a reconnect is not a second vote.
• Never fuses payload fragments from different providers.

Measured behaviour

Each figure below is produced by an example or test in the repository and is reproducible. Slot rate measured on mainnet-beta on 25 October 2026.

Mainnet block sizes were sampled the same day: median 4.26 MiB, largest 7.90 MiB, smallest 2,389 bytes. A round 2 MiB assumption understates the memory ceiling by roughly half.

Using it

[dependencies]
yellowstone-reconcile = { git = "https://github.com/ss1738/yellowstone-reconcile" }

Assign a stable ProviderId per independent provider, call connect for each connection, pass each stream's output with its returned Session to ingest, and halt downstream processing for a slot on Conflict. Advance retention explicitly and handle the pending slots the prune report returns.

Commit to your sink first, keyed idempotently on slot, and only then persist the checkpoint. A crash between those two steps re-delivers exactly one slot, which is why the sink write has to be idempotent.

Status