When translating existing time-series plots to Rerun our logging APIs have very poor ergonomics.
A user must iterate through every sample, calling set_time with the time value, and then logging the individual scalars.
The proposal from:
Would let us optimize this a bit by passing the time values into the log-calls, but will still be order-N which can be quite painful for large time-series.
A direct temporal batch logging API would instead let us collapse this into a single call. For example:
rr.log_temporal_batch(timestamps=[...], ScalarBatch(...))
This would let us dramatically optimize the ingestion by letting us create a pre-batched uniform buffer of scalars that looks like what the eventual output of the batcher is going to be anyways.
When translating existing time-series plots to Rerun our logging APIs have very poor ergonomics.
A user must iterate through every sample, calling set_time with the time value, and then logging the individual scalars.
The proposal from:
Would let us optimize this a bit by passing the time values into the log-calls, but will still be order-N which can be quite painful for large time-series.
A direct temporal batch logging API would instead let us collapse this into a single call. For example:
This would let us dramatically optimize the ingestion by letting us create a pre-batched uniform buffer of scalars that looks like what the eventual output of the batcher is going to be anyways.