Glossary

Terms as Thadam uses them. Where a term has a formal definition in an ADR, the ADR governs.

Ack mode (ack_mode) — When a source message is acknowledged to its broker: on_commit (after the WAL commit; at-least-once) or on_receive (when pulled; at-most-once across a crash).

Admin API — The operational HTTP surface on 127.0.0.1:9464 by default: /healthz, /readyz, /metrics, /status. Reference.

At-least-once — Every accepted message is delivered one or more times; duplicates are possible and expected. Thadam’s default class.

At-most-once — Delivery without retransmission; a crash can lose messages. Applies to QoS 0 and ack_mode: on_receive across a crash.

Batch sink — Sink contract for transactional batch targets (ClickHouse, Kafka, NATS, MQTT publish): deterministic batches retried under the same batch ID. Compare segment sink.

Capture-first WAL — The architectural rule (ADR-0003) that raw messages are appended to a durable write-ahead log before decode, transform, or acknowledgement. See WAL.

Chain — A sink’s private pipeline: WAL reader and cursor → router → decode → dataset shaping → transform steps → sink driver. See per-sink chains.

Claims register — The table in ADR-0011 listing exactly which guarantees Thadam may publish, with conditions and required evidence.

Config hash — A canonical hash of the configuration (algorithm version 1) recorded in thadam.json and every Parquet footer; one of the inputs to replay equivalence.

Cursor — A consumer’s durable position in the WAL (cursors/<name>.cur). Readers and commits never pass the durable end; positions at or below a durable cursor are never reused.

Dataset — An output table: match rule, decoder, identity template, optional event ID/dedup/explode/time rule, and a declared field list. The unit every sink produces. Reference.

Delivery class — What a sink can guarantee: AtMostOnce, AtLeastOnce, or IdempotentBatch (for example effectively-once ClickHouse inserts under probed table settings).

Durable end — The last WAL position covered by a successful fsync; the bound readers and cursor commits respect.

Epoch (WAL epoch) — A random ID created when a data directory initializes; resets WAL positions to 1 and appears in file names and object keys so two epochs never collide.

Event — The canonical typed record produced by decoding one source message (or one exploded sample), with time, quality, identity, and payload fields. Event model.

Event time — The time a sink materializes: corrected time, else source time when the domain is utc, else null.

Field ID — A stable Parquet PARQUET:field_id per column (envelope 1–99, identity 100–199, declared 1000–4,194,303, nested ≥ 8,388,608), never reused. See field IDs.

Gap — An explicit marker in a WAL read sequence — corrupt, evicted, or missing — so loss of position is never silent.

Group commit — Batching WAL appends and fsyncing the batch at most once per fsync.interval; the latency/throughput trade at the heart of the ack window.

Identity — The ordered asset-path hierarchy (site → robot → …) extracted by templates such as {topic.2} and stored as columns.

Ingest time — When the Thadam edge received a message. Always present, never changed.

on_full — The required, explicit policy for a full WAL budget: block or drop_oldest. See disk budgets.

Quarantine — The segment sink for rejected events, with a fixed schema and its own disk budget.

Repair mode (--repair) — A run mode that accepts a thadam.json pipeline-name or WAL-epoch mismatch and rewrites the binding; never accepts newer format versions.

Segment — In the WAL, one append-only file of frames. In sinks, one finalized Parquet (or NDJSON) file for a dataset.

Segment catalog — The append-only, CRC-checked log of Finalized/Eligible/Uploaded/Deleted segment records.

Segment sink — Sink contract for accumulating file targets (Parquet, NDJSON): a watermark reported via poll_commit, finalized files, cursor_committed for ship eligibility. Compare batch sink.

Shipper — The service that uploads eligible segments to object storage; not a sink, so a network outage cannot stall capture.

Source time — The timestamp the device reported, set once at decode and never overwritten.

Time domain — Which clock source_time is on: utc, tai, monotonic, sim, or unknown. Only utc becomes event time.

Warm start — Rebuilding stateful transform state (dedup, on_change, sample) at startup by re-reading [cursor − window, cursor) from the WAL without emitting.

Watermark (sink) — The WAL position through which a segment sink’s files are durable; the runtime commits cursors only up to watermarks.