Recordings exist as a collection of files per node present in the original cluster.

The files for a single node constitute a "timeline". Within those files, multiple sessions exist, corresponding to one client session. Each recorded session contains all observed requests and responses for that specified session, captured immediately after being received/before being sent.

Recordings may become large. We cannot read all requests into memory ahead of time.

Ideally we also should not require that all files be present concurrently. This would allow the files to be incrementally fetched from (e.g.,) S3.