# Archive and grace hold

> How Pulse hands your data to your own bucket on day 7, the file layout and manifest, and what happens when the bucket isn't reachable.

Pulse keeps telemetry for 7 days (UTC days, today included). Then it's yours: every day at 00:30 UTC, Pulse copies the day that just turned 8 days old to your own bucket, checks the copy, and deletes its own.

## Your bucket

Set it in **Settings → Storage** (org owners and admins): any S3-compatible endpoint, such as AWS S3, Cloudflare R2, Backblaze B2 or Google Cloud Storage (interoperability keys). Pulse needs to put, read and delete objects; it checks this when you save. With a key prefix set there, everything below goes under that prefix. The archive is on your storage bill, not Pulse's.

## What lands there

```
pulse/logs/dt=2026-10-04/<id>.parquet
pulse/metrics/dt=2026-10-04/<id>.parquet
pulse/manifests/2026-10-04.json
```

- **Parquet** files (Snappy), one schema per signal, in Hive-style `dt=` folders. Logs: `time`, `observed`, `sev`, `sev_text`, `body`, `service`, `trace_id`, `span_id`, `scope`, `resource`, `attrs`. Metrics: `time`, `start`, `name`, `unit`, `service`, `kind`, `monotonic`, `temporality`, `value`, `count`, `min`, `max`, `buckets`, `resource`, `attrs`. Times are UTC milliseconds; `resource` and `attrs` are JSON. Columns are only ever added at the end.
- The **manifest** is written last, so if it's there the day is complete: `{org, day, schema: 1, files: [{key, signal, rows, bytes, md5, t0, t1}], written_at}`.

Read the files with anything that reads Parquet, for example DuckDB:

```sql
SELECT service, count(*) FROM read_parquet('s3://my-bucket/pulse/logs/dt=2026-10-*/*.parquet') WHERE sev >= 17 GROUP BY 1;
```

The schema is Iceberg-compatible, so PyIceberg's `add_files` can register the files in an Iceberg catalog without rewriting them.

## Grace hold

If no bucket is set, or Pulse can't write to it, the day goes on **grace hold** instead of being deleted:

- Pulse emails the org's owners and admins and sends an [archive alert](/docs/alerts-and-shift/).
- Every day it tries again, and emails a digest of the days held and when they'll be dropped. It warns 48 hours before a drop.
- At the end of the grace period the day is **dropped**: deleted for good. The email carries the day's manifest.

The grace period is 7 days by default; set it from 0 to 30 days on the **Archive** page or with `PUT /api/v1/settings {"grace_days": 14}`. Zero drops a day at once. Held days aren't queryable, and they're billed per GB-day (see [pricing](/pricing)).

Fixed the bucket? Press **Retry now** on the **Archive** page (`POST /api/v1/archive/retry`) rather than waiting for 00:30. `GET /api/v1/archive` and the MCP tool `archive_status` show each day's state: `archived`, `grace`, or `dropped`, with the last error.