TRACKMANIARL / SYSTEM GUIDE / 02
Resume the complete training state
An exact resume restores one continuous run history, including unapplied durable experience.
01 Capture
02 Save durably
03 Recover
Running learner
Replay + optimizer
RNG + counters
Coherent capture
State + applied WAL
frontier
Durable write
Write a temporary file,
fsync, then replace.
Checkpoint 2.0
State, identity and
WAL frontier
Validate recovery
Schema + run identity
Journal + prune floor
Restore + replay WAL
Restore state
Replay WAL by row ID
RECOVERY GATE
A foreign journal or an invalid frontier is
rejected. Prune WAL only after the checkpoint is
durable.
WARM-START IS DIFFERENT
Loading compatible model tensors starts a new
run. It does not restore replay, optimizer or
RNG.
Download editable diagram