InfluxDB 3 sources

tsm2arc reads InfluxDB 3 (3.0–3.11, Parquet engine) object stores directly — parquet data files, snapshot manifests, the catalog, and the WAL — with no running influxdb3 required. Point --datadir at the store and everything is auto-detected:

# local directory (--object-store file, or any store synced to disk):
# point at the store root — the directory holding the node prefix
tsm2arc --datadir /mnt/influxdb3-store --arc-url https://arc.example.net --dry-run

# straight from S3: listings and range reads, no scratch-volume sync.
# Credentials come from the standard AWS chain (env vars, shared config,
# IMDS/IRSA); --s3-endpoint for MinIO and on-prem gateways.
tsm2arc --datadir s3://my-bucket/influxdb3 --arc-url https://arc.example.net --dry-run

No v3-specific flags are needed. Database and table names, series keys, and column types come from the store's catalog, and every catalog format is read natively — the JSON eras of 3.0–3.9 and the binary format of 3.10+/3.11 (decoded by the upstream codec itself, compiled to WebAssembly and embedded in the binary — tsm2arc stays a single static executable). Each live table migrates as one unit, and the whole 1.x/2.x toolbox applies unchanged: dry runs, sizing, resume, verification, --analyze profiling, measurement policies, --db-map.

The WAL is recovered — nothing is left behind

InfluxDB 3 keeps up to ~10 minutes of the newest writes only in its WAL, and a clean server shutdown does not flush them: v3 never snapshots on shutdown, so a cold store almost always carries un-snapshotted WAL — sometimes including entire recently-created tables that exist nowhere else. tsm2arc decodes those WAL files in-process and merges the rows into the migration:

decoded 40 row(s) from 4 un-snapshotted WAL file(s) (sequences 13-16); nothing left behind

--skip-wal skips the decode explicitly. Only use it when you know the WAL content is disposable — the skipped writes will be missing from Arc, and tsm2arc says so loudly.

Correctness semantics

• Duplicates resolve last-write-wins, deterministically: when the same series and timestamp appear in multiple parquet files (or in the WAL), the newest write wins — the same overwrite semantics InfluxDB 3 applies at query time.
• Resume is byte-exact. Emission order is a pure function of the store, and the live file set is part of the checkpoint fingerprint: a store that changed between a run and its resume fails up front as "different settings" instead of corrupting the migration. Migrate from a quiesced or frozen store.
• Three-way verification. The snapshot manifests carry per-file row counts (--analyze prints the reconciled totals per table), tsm2arc reports what it extracted, and Arc can be counted after the load. If the source server can be run, its own SELECT count(*) — which replays its WAL — is the gold oracle; tsm2arc's totals match it exactly.

Enterprise stores

Enterprise clusters keep the catalog under the cluster prefix next to the node prefix; tsm2arc resolves that automatically. Support depends on the compactor:

v3 flags reference

All optional — the common case needs none of them:

--s3-endpoint URL     custom S3 endpoint (MinIO/on-prem); forces path-style addressing
--s3-region REGION    override the AWS region from the credential chain
--skip-wal            skip decoding un-snapshotted WAL (those newest writes will
                      be MISSING from Arc; normally not needed)
--v3-db id=name                     override a database id's name
--v3-table db/table=name:tag1,tag2  override a table's name and series key
                                    (tags in first-write order)

Troubleshooting