Status group¶
Execution lifecycle for a record is stored in the status group. It is used
for monitoring, resume decisions, and UI summaries. Scientific result
arrays live on frame, trajectory, or observables; time series of
measurements live on metrics.
In the reference binding the contents are group attributes (one JSON
object). status.state is required whenever the group exists.
status
+-- state: string[]
+-- (stage: string[])
+-- (epoch)
+-- (global_step)
+-- (message: string[])
+-- (started_at: string[])
+-- (updated_at: string[])
+-- (finished_at: string[])
+-- (error)
\-- ...
state
Lowercase lifecycle: pending, running, succeeded, failed,
cancelled, skipped. Writers may preserve custom states.
state is what the writer last recorded, not a liveness signal: a process
that is killed leaves running behind for ever. A reader that must tell a
live run from a dead one compares updated_at with the cadence the host
expects, or asks the host that started the run; it never takes running as
permission to treat the record as locked, nor its absence as permission to
append (Chunking and packing).
stage
Current execution phase, independent of lifecycle state (e.g. train,
eval, simulate).
epoch, global_step
Reserved counters: non-negative JSON integers. Extra counters go under
status.progress, an object of named numbers.
message
Free text for a human.
started_at, updated_at, finished_at
RFC 3339 timestamps with an
explicit offset (Z or ±hh:mm); a timestamp without one names no
instant and is not valid here.
error
What went wrong: an object with at least message (string), and optionally
type (string) and traceback (string). A writer that records failed
SHOULD record error with it; a failed status without one is valid
(a run killed from outside has nothing to say), and so is an error kept
beside a later state (a retry that succeeded).
Every key a producer adds beyond these is preserved.
A run-shaped record is meta plus status, optionally with metrics and
method. frame is optional. See Run surface.