Export

Export writes queryable pond data as external Parquet files. It does not create a replica, backup attachment, capsule, or restoreable copy of all pond metadata. Keep authoritative source data and replication/recovery procedures separately.

pond export --pattern '/sensors/*' --dir OUTPUT accepts repeated patterns; quote globs to prevent shell expansion. Inputs must be queryable tables/series, not arbitrary raw file bytes. A format URL supplies the decoder for raw input. An output directory is a host side effect, not an in-pond path.

CLI and sitegen are separate contracts

The CLI reference describes --temporal, --start-time, --end-time, and --rebuild. CLI temporal levels are comma-separated year/month/day/hour/minute names. CLI time filtering uses the timestamp column with inclusive start/end comparisons; do not assume a configurable timestamp column just because sitegen has one.

Sitegen export stages instead specify name, pattern, target_points, and timestamp_column. They discover resolutions and choose Hive-style partition levels so a viewport needs few files. target_points is sizing guidance, not sampling/aggregation itself. Template routes use the export's wildcard captures and file manifest.

Time, partitions, and schemas

Parquet contains typed Arrow data. Time partition directory labels describe the partition interval, not a promise that both endpoint samples exist. Different sources can have different timestamp units/timezones; inspect schema and exported values. Filter boundaries and aggregation bucket boundaries are separate concepts. Do not infer the result from filename count alone.

The timeseries site package asserts page URLs, the referenced Parquet manifest, and exported source row counts/values. Failures can leave partial output; publish a fresh completed directory. --rebuild drops derived reduction cache state for recomputation; it is not a data recovery or source mutation command.