# transformers CI telemetry

Public, daily-partitioned snapshot of the transformers CI test telemetry
collected by the pytest observability stack (OpenTelemetry → Tempo). Refreshed
**hourly**. Schema version **v1**.

This is derived from raw test-execution traces so you can build apps and
analyses on top of CI data without access to the internal stack.

## Layout

```
current_view.json          manifest: schema_version, updated_at, partitions, totals
README.md                  this data card
daily/
  <YYYY-MM-DD>/            partition = UTC day of the run's start
    test_rows.parquet       one row per (trace_id, test_nodeid)
    run_rollups.parquet     one row per (run_id, test_job)
    traces/
      <trace_id>.json       raw Jaeger-shaped trace (full fidelity)
```

The bucket is the long-term archive; it keeps full history independent of the
stack's operational retention.

## `test_rows` columns

- `ts`
- `date`
- `service_name`
- `provider`
- `pr`
- `run_id`
- `trace_id`
- `test_job`
- `test_nodeid`
- `test_module`
- `test_class`
- `test_function`
- `test_line`
- `model`
- `gpu`
- `status_code`
- `duration_seconds`
- `exception_type`
- `exception_message`
- `exception_stacktrace`

`model` is derived from `tests/models/<model>/...` nodeids (empty otherwise).
`gpu` is derived from the job name (`single` / `multi` / empty).
`status_code` is the OTEL span status: `OK` / `ERROR` / `UNSET`.
`exception_message` and `exception_stacktrace` are the **full, untruncated**
failure text.

## `run_rollups` columns

- `date`
- `service_name`
- `provider`
- `pr`
- `run_id`
- `test_job`
- `total_tests`
- `passed_tests`
- `failed_tests`
- `duration_seconds`
- `start_time`
- `end_time`
- `job_count`
- `commit_sha`
- `commit_message`

`job_count` is the number of distinct jobs that contributed tests to the run.

## Load examples

```python
import pandas as pd
df = pd.read_parquet(
    "hf://buckets/huggingface/transformers-ci-telemetry/daily/2026-06-08/test_rows.parquet"
)
```

```sql
-- DuckDB, straight from the bucket
SELECT model, count(*) FILTER (WHERE status_code = 'ERROR') AS fails
FROM 'hf://buckets/huggingface/transformers-ci-telemetry/daily/*/test_rows.parquet'
GROUP BY model ORDER BY fails DESC;
```

## Notes

- Coverage reflects what the stack actually traced — uninstrumented workflows
  are absent here too.
- Columns are additive across schema versions; `current_view.json.schema_version`
  is bumped on any change.