While preparing for the Sei Giga upgrade, Sei Labs benchmarked node performance using different state store backends. When comparing RocksDB and PebbleDB with MVCC for indexing-heavy historical queries, RocksDB was able to reduce node operator latency by 10–40×

We measured traceBlock latency across several million blocks — an RPC query that performs extensive iteration over key/value pairs in the state store. As node history increases, PebbleDB iteration time grows dramatically, while RocksDB maintains a much flatter latency curve.

The primary reason for this difference lies in backend design. RocksDB supports native user-defined timestamps for MVCC and efficient column families. PebbleDB, by contrast, lacks native versioning, requiring key suffixing and manual iteration over multiple versions.

This leads to significantly better iteration efficiency in RocksDB — especially for nodes with large historical state. As stored history increases, the performance gap between PebbleDB and RocksDB becomes more pronounced.

While RocksDB introduces a minor build tradeoff, the performance improvements make it a strong option for archive and RPC nodes handling large state or long history. We've observed consistent improvements in trace latency and iteration speed across setups