Seven brokers. Nothing in production moves.
Kafka, RabbitMQ, NATS, AMQP 1.0, Redis, MQTT and STOMP, each an item in the project with the same log, composer and presets. The rule underneath them all: looking is not taking. A read in Sonda never advances a consumer, never acks on your behalf, never moves a message a production system is waiting for.
Looking is not taking
Looking is not taking.
A Kafka read assigns the partition directly and starts where you say — the end, the start, an offset, a point in time — without joining a consumer group or committing an offset, so the production consumer's position is exactly where it was. A RabbitMQ message in HOLD is seen in the log and never taken; it goes back to the queue when you stop. A NATS stream is read through an ephemeral ordered consumer that never acks. AMQP 1.0 and Redis leave the settling to you, one message at a time.
Pure Go everywhere — no native client libraries, nothing to install beside Sonda. The Network page's CAs and client certificates apply to every one of them.
What each one does, in plain words.
Seven items of the project, one log, one composer, one set of presets — and each broker's own verbs where it has them.
Kafka: read any partition from any point, produce and read it back.
A bootstrap list, TLS, SASL PLAIN and SCRAM-SHA-256 / 512. Topics with their partitions, leaders and log offsets. Read one partition or all, from the end, the start, an offset or a point in time. Produce with a key, headers and a chosen partition; the answer says partition@offset and offers to read again from there. The cluster is probed every 15 s so the pane is never wrong about being connected.
RabbitMQ: consume the way you mean to, and manage without the plugin.
AMQP 0-9-1 with the channel in confirm mode. Consume with auto-ack, ack once seen, HOLD, or reject to drop or dead-letter; prefetch and exclusive. Peek or take one message. Publish with every property and header; the mandatory flag shows unroutable messages as returned. Reply on a message's reply-to. x-death dead-letter history decoded.
With the management API: overview, queues, exchanges and bindings, refreshed every 5 s. Without it: declare, bind, purge and delete over AMQP itself.
NATS: core, JetStream and Key/Value, from one item.
A server or a cluster list; user and password, token, NKey seed or a .creds file. Subscribe with wildcards and queue groups. Publish, or request and wait for the first answer, timed. JetStream: publish with ack and dedupe, list streams and consumers, read from new, all, last, last-per-subject, a sequence or a time. Key/Value: buckets, keys, get, history, put with a revision, delete, purge, watch a bucket or a key pattern.
AMQP 1.0: Service Bus, Event Hubs, Artemis, IBM MQ and the rest.
For Azure Service Bus and Event Hubs, ActiveMQ Artemis and Classic, Qpid, Solace, IBM MQ, RabbitMQ 4. Paste an Azure connection string and it fills in. SASL anonymous, plain or external with a client certificate. Receive with a credit, a JMS selector, an Event Hubs partition, a Service Bus session — or the next free one. Request mode with a temporary reply address. Typed application properties, annotations, a scheduled enqueue time, the body as bytes or as a JMS text message.
Redis: streams with their groups, and Pub/Sub.
Standalone or cluster, TLS, an ACL user, a database, a client name. Streams found on every master with their stats, consumer groups with pending entries; read with XREAD from any point or as a consumer with XREADGROUP; ack from the log; first and last hundred; add with ordered fields and a trim; create or destroy a group. Pub/Sub with patterns, and how many subscribers a publish reached.
MQTT and STOMP: the device side and the classic brokers.
MQTT 5 and 3.1.1 over tcp, tls, ws and wss; QoS 0, 1 and 2; wildcards; the retained flag explained; MQTT 5 properties; reply on a response topic. STOMP 1.0 to 1.2 for ActiveMQ, Artemis, the RabbitMQ STOMP plugin and Apollo, over TCP, TLS or WebSocket, with selectors, prefetch, durable subscriptions, heart-beats and a timed receipt.
Produce, read it back, and leave production alone.
Beside it, a Kafka pane: the topic's partitions with their leaders and offsets, a read of one partition from a point in time, and a produce whose answer says where the message landed and offers to read again from there. The production consumer's committed offset is on the same screen — and it has not moved.
The same pane, the same verbs, on every broker: the log follows, pauses when you scroll back, and filters to one topic or subject.
A message is an event. An event can start work.
Every broker here can start a flow when a message matches a topic and a filter — and the flow's outcome settles the message.
A flow listens
On a topic with wildcards and a filter expression, from the same connection the pane uses, with the same permissions.
A message matches
The flow runs, N at once, up to a queue of 500. A RabbitMQ reply-to is answered by the flow.
The outcome settles it
A run that ends well acks; on failure the message is rejected or requeued, as the trigger says.
A breaker watches
30 runs in 10 seconds switch the flow off, so a storm on a topic never becomes a storm of calls.