arcli delete
Delete rows from an Arc measurement by SQL predicate with arcli delete --where: dry-run preflight, confirmation and --yes, the delete.enabled gate, NULL-predicate handling, and time-bounded deletes.
arcli delete removes the rows of one measurement that match a SQL predicate, by rewriting the Parquet files that contain them. It needs an admin token and a server with delete.enabled = true in arc.toml; in a cluster only the primary writer accepts it.
Quick reference
arcli delete --database metrics --measurement cpu --where "host = 'web-4'" --dry-run
arcli delete --database metrics --measurement cpu --where "host = 'web-4'"
arcli delete --database metrics --measurement cpu --where "time < '2026-01-01' AND region = 'eu-west'" --yes$ arcli delete --database metrics --measurement cpu --where "host = 'web-4'" --dry-run
Would delete 1 rows from metrics.cpu across 1 file(s)
duration: 2ms
$ arcli delete --database metrics --measurement cpu --where "host = 'web-4'"
Delete 1 rows from metrics.cpu by rewriting 1 file(s)? Server limits: confirmation threshold 10000 rows, max 1000000 rows per delete. [y/N] y
Deleted 1 rows from metrics.cpu (1 file(s) rewritten, 1 affected)
duration: 10ms| Flag | Description | Default |
|---|---|---|
--database string | database holding the measurement | |
--dry-run | report what would be deleted without deleting | |
--measurement string | measurement to delete from | |
-o, --output string | output format: table|json | "table" |
--where string | SQL predicate selecting the rows to delete ("1=1" for all) | |
-y, --yes | skip the preflight scan and the confirmation prompt |
Plus the connection flags.
How the predicate is applied
--where is a SQL boolean expression over the measurement's columns. arcli sends it as (<where>) IS TRUE, so rows where the predicate evaluates to NULL (a missing tag, a NULL field) are kept, the same way a WHERE clause keeps them out of a result. Put time bounds inside the predicate; there is no separate --before or --after.
Before deleting, arcli runs the same request as a dry run and folds its row and file counts, together with the server's limits (the row count above which Arc itself demands confirmation, and the maximum rows per delete), into the [y/N] prompt on stderr. --yes skips both the preflight and the prompt, so keep it for scripts whose predicate you have already tested with --dry-run. When the preflight reports zero matching rows, arcli additionally checks the predicate with a COUNT, so a typo in a column name is an error rather than a silent "0 rows".
Runtime and interruption
A delete rewrites every file that holds a matching row, so it can take minutes on a large measurement; the default --timeout is 30m. If arcli times out or you press Ctrl-C, the server continues the rewrite: check with a query rather than re-running blindly. In the output, affected is the number of files that contained matching rows and rewritten the number the server rewrote successfully.
For age-based deletion on a schedule, use retention policies instead.
arcli cq
Arc continuous queries from arcli: cq create, list, show, update, delete, execute and executions, the {start_time}/{end_time} placeholders, database-qualified sources, and windowed manual runs.
arcli backup
Server-side Arc backups from arcli: backup create, list, show, status, restore and delete, the backup.local_path target, --wait polling, the single operation slot, and what a restore overwrites.