Statistics and Costs

Statistics and Costs

stats = dib.summarize(out)                 # out = the result of run()
print(stats)
# 10000 rows: 9987 ok, 13 failed (99.9% success); 412 reused; tokens 2140311 in / 381220 out;
# 37 retries; latency p50 812 ms, p95 2310 ms; errors: SchemaValidationError x9, RateLimitError x4

summarize(output, output_col="output") works on whatever run() returned: a list of dicts, a pandas DataFrame or a Spark DataFrame. It also works on a table you saved earlier. It reads only the <out>_error and <out>_meta columns. On Spark it is one aggregation plus one small group-by, and no rows are collected. For a lazy Spark result (no checkpoint) it triggers inference, so summarize what you wrote, not the lazy DataFrame.

RunStats

field	meaning
rows, succeeded, failed	row counts; success_rate = succeeded / rows
reused	successful rows that were not sent to the model (attempts == 0: dedupe copies)
retries	API attempts beyond the first, summed over rows
prompt_tokens, completion_tokens	token usage reported by engines
audio_seconds	total transcribed audio
latency_ms_p50, latency_ms_p95	per-row latency percentiles (approximate on Spark)
errors	{error class: count}

to_dict() returns all fields, and str(stats) gives a one-line summary.

Local and multiprocessing runs also log the summary and wall time on the distribatch logger (logging.getLogger("distribatch"), level INFO).

Costs

distribatch ships no price tables, because prices change. Pass yours:

stats.cost(input_per_million=0.50, output_per_million=1.50)   # tokens
stats.cost(per_audio_minute=0.006)                             # transcription

Embedding requests with several inputs report usage per request, not per row. For those runs prompt_tokens undercounts, so use the provider's bill.

MLflow

with mlflow.start_run(run_name="ticket-routing-2026-10"):
    out = job.run(df, backend="spark", checkpoint=..., id_col="ticket_id")
    dib.summarize(out).log_to_mlflow(job)

This logs every numeric field as distribatch.<field>, error counts as distribatch.errors.<Class>, and the job's task, model, engine and fingerprint as params. Needs distribatch[mlflow].