FAQ

Is Thadam production-ready?

No — it is in beta. The architecture is settled, the core crates are tested, and thadam run executes a pipeline end to end. The WAL’s on-disk format is not frozen yet: it freezes at the v0.1.0 tag, after the ADR-0004 re-run on reference hardware. See the status banner and roadmap.

Does Thadam guarantee exactly-once delivery?

No, and it will not claim to. Exactly-once is a sink-specific capability, not a universal property (product.md section 27). Thadam’s defaults are at-least-once capture plus durability; “effectively-once” to a sink is possible only with specific downstream settings (for example a replicated ClickHouse table or a deduplication window), and Thadam probes and reports the effective class rather than assuming it. The binding wording is the claims register.

What happens when the disk fills?

Whatever your wal.on_full says — and there is deliberately no default. block stalls sources and moves loss into the broker’s queue; drop_oldest evicts old segments with the loss counted and surfaced as gaps. This is the only place inside Thadam where acknowledged data can be lost. Details: backpressure and disk budgets.

What happens when the network is down for hours?

Capture continues into the local WAL within its disk budget; delivery stalls and retries with backoff. When connectivity returns, the WAL drains and uploads resume. Dimension the budget for the longest outage you need to absorb (retention sizing) and size the broker’s session queue for the downtime when Thadam is the side that is down.

What throughput can I expect?

Per connection, roughly broker inflight window ÷ (commit latency + RTT) with ack_mode: on_commit — about 200–320 msg/s with default broker windows at a 100 ms commit, about 1,700 msg/s with Mosquitto’s default window at the provisional 10 ms group commit (MQTT spike; measured on one laptop, ratios transfer, absolute numbers need your environment). The throughput ceiling page lists the levers. Published cross-architecture benchmark results arrive with the v0.1 reliability gate; nothing else is claimed.

Why another data pipeline instead of Kafka Connect / NiFi / a vendor gateway?

Those are strong systems; the gap Thadam aims at is the combination, in one small Rust binary: robotics-aware semantics (time domains, quality, identity as data), edge-native reliability (durable local buffer, replay, explicit loss behavior), and modern analytical sinks (Parquet/object storage, ClickHouse). Thadam complements Kafka/Flink-style backbones rather than replacing them (product.md section 34).

Does Thadam control my robot or carry commands?

No. Thadam may transport command/event-like messages only where explicitly configured, and it does not present itself as safety-critical control infrastructure. Motion planning, control loops, and safety systems are out of scope (what Thadam is not).

Which protocols and sinks work today?

Sources: MQTT (3.1.1; v5 experimental), HTTP, gRPC, Modbus TCP, SocketCAN (Linux only), OPC UA (Cargo feature opcua), ROS 2 MCAP (offline ingestion), and live ROS 2 over rmw_zenoh (Cargo feature ros2-live, experimental), with JSON, ROS 2 CDR, and Sparkplug B decoding. Sinks: Parquet (local segments), S3-compatible object storage shipping (thadam-object-store), Apache Iceberg table registration (Cargo feature iceberg, experimental, P2), NATS JetStream, Apache Kafka, MQTT publish, and the ClickHouse batch sink (P1). The roadmap has the full sequence.

Can I run Thadam on a Raspberry Pi / Jetson / Mac?

Linux ARM64 and x86_64 are supported and are the only durability targets, on industrial SSD or eMMC that honors flush; SD cards are not a supported target for durability claims. macOS builds work for development but carry no durability claims. Reference-hardware validation (x86 NVMe, ARM eMMC, a Jetson/Pi class board) is an open readiness item.

How do I keep credentials out of the config file?

${env:NAME} or ${file:/absolute/path} interpolation (secrets): values are redacted from logs and diagnostics, hashed as written so rotation does not churn the config hash, and the MQTT password specifically uses password_file. Container secret mounts (/run/secrets/…) work well, as the demo shows.

Where is the Python SDK?

The SDK is implemented in sdk/python/ (PyO3 and maturin) and documented in the Python SDK page; build it from a checkout with maturin develop. It is not published to any package registry yet: its package version is 0.1.0b1 (Cargo 0.1.0-beta.1).