Historical data format is the same as provided by real-time Coinbase Pro WebSocket API with addition of local timestamps. If you'd like to work with normalized data format instead (same format for each exchange) see downloadable CSV files or official client libs that perform data normalization client-side.
Tardis-machine is a locally runnable server that exposes API allowing efficiently requesting historical market data for whole time periods in contrast to HTTP API that provides data only in minute by minute slices.
During recording, data integrity of order book L2 incremental updates messages is being validated using sequence numbers provided by Coinbase real-time feed - in case of detecting missed message WebSocket connection is being restarted.
full_snapshot - generated channel with full order book L3 snapshots
Coinbase Pro (formerly GDAX) real-time WebSocket API provides initial full order book snapshots for level2, but not for full channel. To overcome this issue we fetch initial order book snapshots from REST API and store them together with rest of the WebSocket messages. Such snapshot messages are marked with "type":"full_snapshot" and "generated":true fields. This effectively allows reconstructing historical full order book for full (L3) channel. Validation based on sequence numbers if full_snapshot overlap with WS L3 updates has been added 2020-06-11.