mq-bridge library

Docs.rs

      ┌────── mq-bridge-lib ──────┐
──────┴───────────────────────────┴──────
            crossing streams

mq-bridge is an asynchronous message library for Rust. It connects message brokers, databases, files, HTTP/WebSocket endpoints, and in-memory channels behind one small set of traits.

It is not only a forwarder. A route can transform, filter, fan out, retry, rate-limit, deduplicate, or turn a request into a response before the message reaches the next system. The core is built on Tokio and keeps the transport details at the edge, so application code can mostly work with CanonicalMessages and handlers.

Move data from A to B — inside your own service

If you need to move data or events reliably between systems and you write code (Rust, Python, or Node), mq-bridge is a strong default. It is a library you embed, not a daemon or control plane you operate.

Prefer not to write code? mq-bridge-app runs the exact same engine as a standalone, zero-code ETL service configured entirely by YAML or environment variables — move data from A to B without writing a line. It ships a Postman-style UI to build, send, and inspect messages against a route, and can import Postman collections and AsyncAPI documents to scaffold routes and endpoints for you.

• 16+ transports, one API: Kafka, NATS, AMQP (RabbitMQ), MQTT, MongoDB, Postgres CDC (logical replication), PostgreSQL / MySQL / SQLite (SQLx), ClickHouse, HTTP, WebSocket, gRPC, ZeroMQ, Redis Streams, AWS SQS/SNS, cloud object storage (S3 / GCS / Azure), IBM MQ, files, dir_spool (a crash-safe directory FIFO queue), and in-memory channels — all behind the same receive_batch / send_batch shape.
• Change Data Capture: stream row-level changes from Postgres (logical replication / pgoutput) and MongoDB (change streams) as flat rows with an operation marker.
• Restart-safe delivery: batch-aware ack/nack with commit sequencing for cumulative-ack brokers; the integration suite shows no data loss during in-flight broker restarts, including a Postgres CDC restart-safety test.
• Reliability middleware, not a framework: retries, dead-letter queues, deduplication, rate limiting, and cookie/session persistence wrap any endpoint.
• Polyglot on one engine: the same Rust core ships as native Python and Node.js bindings; routing, batching, and broker I/O stay in Rust.
• TLS everywhere, one config shape: a single TlsConfig block (CA bundle, client cert/key for mTLS, insecure-skip) is reused across transports.
• Self-hosted, no daemon: generate config in the optional UI, paste it into your code, run it in-process. No hosted control plane, no separate scheduler.

Throughput & footprint. In our own benchmarks, the same engine — driven the zero-code way through mq-bridge-app — on a CSV→JSONL file conversion (1,000,000 mixed-type rows, ~116 MiB) sustained 2,824,858 rows/s at ~28 MiB (mq-bridge 0.4.12), about ~145x faster and ~16x leaner in memory than Meltano (tap-csv → target-jsonl, ~19,500 rows/s / ~444 MiB, measured in an earlier session on the same machine). On the same file in the same session it also ran ~1.4x faster than DuckDB using all of this machine's cores (2,036,659 rows/s) while holding ~17x less memory — DuckDB being a throughput ceiling for the conversion itself, not an ETL tool. Methodology and reporting rules are in benches/ETL_BENCHMARKS.md; the measured numbers and their baselines are in the ETL benchmark harness.

Kafka → file. In a 1,000,000-row relay with the default file format and no transform, the engine was ~65% faster than Sea Streamer comparing both on the mimalloc allocator (~80% against its default-allocator build). The mq-bridge-app benchmark contains the reproducible helper and native file-format caveats.

Language Bindings

mq-bridge is a Rust library, but the same engine ships as native bindings for Python and Node.js. The Tokio runtime, broker I/O, routing, and batching all stay in Rust; the binding is a thin layer for handlers and configuration.

The constructor names are kept aligned across languages, so a config loader reads the same in either binding (Python uses snake_case, Node uses camelCase):

• Route.from_file / Route.fromFile — load a route from a YAML/JSON file
• Route.from_str / Route.fromStr — load from an in-memory YAML/JSON string
• Route.from_config / Route.fromConfig — load from a dict / JS object
• the matching Publisher.* constructors build a publisher endpoint

The name argument is optional in both: pass it to select one entry from a routes:/publishers: document, or omit it to treat the config as a single bare route/endpoint body.

The Python binding also holds up well under load on the third-party http-arena.com requests-per-second HTTP benchmark (live leaderboard — rankings shift over time). See the Python analysis notes for the local HTTP comparison harness.

Architecture

See ARCHITECTURE.md for a detailed overview of the internal design, extensibility, and usage patterns.

Usage Types:

1. Event Handler (TypedHandler): Communicate between applications using strongly-typed message handlers, optionally with response support.
2. Compute Handler: Generally receive and process messages with a custom handler
3. Direct Endpoint Usage: Use publish / publish_batch and receive / receive_batch directly on endpoints. This mode requires manual commit, batch sequencing, and concurrency handling.

For implementation details and quick start examples for each usage type, see the Architecture Guide.

Features

• Supported Backends: Kafka, NATS, AMQP (RabbitMQ), MQTT, MongoDB, SQL Databases (PostgreSQL, MySQL, SQLite via sqlx), Postgres CDC (logical replication), ClickHouse, gRPC, HTTP, WebSocket, ZeroMQ, Redis Streams, Files, dir_spool (a crash-safe directory FIFO queue), cloud object storage (S3 / GCS / Azure), AWS (SQS/SNS), IBM MQ, and in-memory channels.
• Change Data Capture: PostgreSQL logical replication (postgres_cdc) and MongoDB change streams, surfaced as flat rows with an *.operation marker (insert/update/delete/truncate).

  Note: IBM MQ is included in the full feature set via the ibm-mq feature, which loads the IBM MQ client library at runtime via dlopen — no IBM SDK is needed to build. The IBM MQ redistributable client only has to be present at runtime, and only if you actually use an IBM MQ endpoint (it is loaded lazily on first connect). If the client is missing, the affected route fails fast with a non-retryable error instead of reconnecting forever. The loader finds the client via the platform default name, MQ_INSTALLATION_PATH (e.g. /opt/mqm), or an explicit MQB_IBM_MQ_LIB path. To link the client statically at build time instead, use the ibm-mq-static feature (requires the IBM MQ SDK). See the mqi crate for details.

  TLS: IBM MQ has its own TLS config shape (IbmTlsConfig), since the native client doesn't consume PEM files. The field names mirror the generic TlsConfig for config parity, but carry MQ-native semantics: tls.cert_file (alias key_repository) is a CMS key repository path (e.g. /path/to/tls for tls.kdb), not a PEM file. The repository can either be passwordless (backed by a .sth stash file next to the .kdb) or password-protected via tls.cert_password (alias key_repository_password; requires an IBM MQ client/server at 9.3.0.0 or later, which is the capability level ibm-mq/ibm-mq-static build against; 9.2 is EOL).

• Configuration: Routes can be defined via YAML, JSON or environment variables.
• Programmable Logic: Inject custom Rust handlers to transform or filter messages in-flight.
• Batching: Every endpoint uses the same send_batch / receive_batch shape. Routes default to batches of up to 512 messages, configurable with batch_size.
• Middleware:
  • Retries: Exponential backoff for transient failures.
  • Dead-Letter Queues (DLQ): Redirect failed messages.
  • Deduplication: Message deduplication using sled.
  • Limiter: Best-effort throughput limiting in messages per second, including batch-aware pacing.
  • Cookie Jar: Persist and re-inject HTTP-style cookies or selected metadata across requests.
• Concurrency: Configurable concurrency per route using Tokio.

Philosophy & Focus

The project has one main bias: move data reliably without forcing the rest of the application to care too much about the transport.

That means mq-bridge tries to keep the boring parts boring. Kafka offsets, RabbitMQ nacks, HTTP responses, MongoDB polling, WebSocket frames, and file rows are all different in real life, but route code should still be able to receive a batch, process it, publish it, and commit it.

Batching is a big part of that design. Every endpoint is optimized around batch-shaped APIs, even when the backend itself only has a single-message primitive. Routes default to opportunistic batches of up to 512 messages: the consumer waits for the first message, then takes whatever else is already available without adding idle latency. Lower batch_size when tighter failure isolation or a smaller per-batch memory footprint matters.

The error handling follows the same idea. Batch publishing can report partial success, retryable failures, and non-retryable failures. Route commits are sequenced for cumulative-ack brokers so later batches cannot acknowledge earlier unresolved messages; transports with independent acknowledgements can commit batches concurrently. In other words: batching is not just a performance trick bolted onto the side; ack/nack behavior and retry/DLQ handling were built to work with it.

What it does not try to be: a domain framework, an actor runtime, or a full stream processor. You can build CQRS-ish flows with it, but the library cares more about transport, routing, and delivery behavior than about prescribing your domain model.

How it compares

mq-bridge overlaps with config-driven ETL / pipeline runners, but comes in two shapes: a library you embed in your own service, or mq-bridge-app — the same engine as a zero-code, config-driven ETL service. You get the developer-first path and the no-code path from one codebase.

• Choose the mq-bridge library when you write the code that moves the data — in Rust, Python, or Node — and want batching, retries/DLQ/dedup, request-reply, CDC, and TLS behind one small API you run in-process. No separate scheduler, no daemon to operate.
• Choose mq-bridge-app for zero-code ETL when you'd rather move data A→B purely by config: define routes and endpoints in YAML/env, then use its Postman-style UI to send test messages and watch them flow — and import Postman collections or AsyncAPI specs through the UI to generate the config. Same reliability engine, no build step.
• Choose a dedicated stream processor when you need windowed joins or aggregations over time.

Either way it stays transport-first on purpose: it moves and routes data reliably rather than prescribing a domain model or a platform.

Status

mq-bridge is young (created in 2025) but its reliability behavior is exercised by an automated integration + performance suite across every supported endpoint:

• Integration and performance tests cover all endpoints in both queue and subscriber modes, request-reply where supported, and protocol-specific behavior.
• All endpoints have automated integration tests that showed no data loss during in-flight broker restarts; MQTT publish-confirmation was hardened until a chaos test drove in-flight loss to zero.
• Postgres CDC ships with a restart-safety integration test: an un-acked, in-flight batch is redelivered after a database restart with no loss and no gap into already-acknowledged changes.
• Throughput is tracked continuously on a public benchmark dashboard; ETL/CDC-shaped, like-for-like scenarios and their fixed methodology are documented in benches/ETL_BENCHMARKS.md.

Known rough edges (be deliberate here):

• old or very new broker-server versions, or unusual broker settings
• subscribe/event and response patterns where the backend has no native equivalent (emulated)
• NATS without JetStream
• TLS: one unified TlsConfig shape is reused across transports and exercised by automated tests; per-backend certificate matrices (self-signed / CA-signed / mTLS / expired-cert rejection) are being expanded — treat non-trivial TLS setups as needing your own verification for now.

Test Notes

• NATS: Automated tests are only run with JetStream enabled. Other NATS modes are not covered by integration tests.
• MongoDB: The reply pattern was only tested in an automated test and is not yet used in projects; because it uses emulation that wait for messages, it may cause severe issues if timeouts are not configured correctly.
• Performance Tests: These are generally executed in non-subscriber (queue) mode for all endpoints.
• Request-Reply: Only tested for endpoints that natively support or emulate it (see backend table below for details). Endpoints like SQLx, Files, AWS, IBM MQ, and Sled do not support request-reply and are not tested for this pattern.
• Subscriber Mode: You may also completely emulate a subscriber mode, if the subscribers are static, by performing a fanout and manually create an endpoint for each target.

When to use mq-bridge

• Hybrid Messaging: Connect systems speaking different protocols (e.g., MQTT to Kafka) without writing a custom adapter for every pair.
• Batch-heavy Pipelines: Increase throughput by moving messages in batches while keeping per-message ack/nack decisions.
• Infrastructure Abstraction: Write business logic against CanonicalMessages and swap the underlying transport later.
• Resilient Pipelines: Apply retry, DLQ, deduplication, limiter, and cookie/session behavior consistently around endpoints.
• Database Integration: Combine databases with message brokers, for example by ingesting messages into SQL/MongoDB or forwarding outbox rows to a broker.
• Sidecar / Gateway: Run the bridge beside another service to ingest, filter, and route messages before they reach the core application.

When NOT to use mq-bridge

• Stateful Stream Processing: For windowing, joins, or complex aggregations over time, dedicated stream processing engines are more suitable.
• Domain Aggregate Management: If you need a framework to manage the lifecycle, versioning, and replay of domain aggregates (Event Sourcing), use a specialized library. mq-bridge handles the bus, not the entity.
• Protocol-Specific Power Features: mq-bridge intentionally exposes a common subset: publish/consume, pub/sub where possible, request-reply where possible, batching, middleware, and ack/nack handling. If your application depends on highly specific broker features, using that broker's native client directly may be better.

Core Concepts

• Route: A named data pipeline that defines a flow from one input to one output.
• Endpoint: A source or sink for messages.
• Middleware: Components that intercept and process messages (e.g., for error handling).
• Handler: A programmatic component for business logic, such as transforming/consuming messages (CommandHandler) or subscribe them (EventHandler).

REFERENCE.md lists every middleware and every structural endpoint (ref, fanout, switch, request, response, reader, static, stream_buffer, null, custom) with fields, defaults and working examples.

To add an endpoint of your own, see EXTENDING.md — or PLUGINS.md to ship a Rust endpoint or middleware as a native library that Rust, Python and Node.js hosts all load at runtime, without compiling it into mq-bridge.

Backend Features & Configuration

mq-bridge endpoints generally default to a Consumer pattern (Queue), where messages are persisted and distributed among workers. To achieve Subscriber (Pub/Sub) behavior, specific configuration is required.

The table below summarizes the capabilities and configuration for each backend:

Plugin Endpoints

These endpoints live in their own repositories and load as plugins, so the core library carries none of their dependencies:

Outside mqb (Python, Node.js, your own Rust host), or for Connect, install the plugin with brew install marcomq/tap/<repo> or conda install -c marcomq <repo>; it is then discovered automatically.

Feature Details

• Request-Reply:
  • Native: Uses protocol-level correlation (e.g., HTTP connection, NATS reply subject).
  • Emulated: Publishes a new message to a reply destination (specified by the reply_to metadata field) carrying a correlation_id metadata field.
• Nack Support: If "Yes", the backend supports explicit negative acknowledgement triggering redelivery. "Eventual" means redelivery depends on timeout or connection drop. "Simulated" is handled in-memory by the bridge. "Consumer-local" means the nacked message is retried inside the consumer process and is lost if that process dies — the producer is never notified.

Database Sources: Change Capture vs Polling

Databases have no native pub/sub, so mq-bridge reads them as a source in one of two ways:

• Change Data Capture (CDC) — tails the database's own change log, capturing inserts and updates/deletes and resuming from a durable log position.
  • On PostgreSQL, the postgres_cdc endpoint streams a logical-replication slot (pgoutput). It emits flat JSON rows tagged with postgres.operation metadata (insert/update/delete/truncate). Acking a batch confirms the LSN back to the server (standby_status_update); a nack (or an interrupted run) does not advance the confirmed LSN, so replication resumes from the last acknowledged position on reconnect — at-least-once, verified by a restart-safety integration test. Enable with the postgres-cdc feature; requires wal_level = logical and a publication. See below.
  • On MongoDB, set consume: capture_new to watch an existing collection for changes from now on, or consume: capture_all to read the existing documents first and then keep capturing changes (no gap, at-least-once). It emits insert/update/replace/delete events tagged with mongodb.operation metadata and checkpoints progress under cursor_id. Both use the change stream (full post-image via updateLookup) and therefore need a replica set — a single-node one is enough. Without one they refuse to start; use consume: snapshot for a one-shot non-destructive read, or consume: consumer for a work queue.
• Cursor polling — pages an existing table by a monotonic cursor_column (WHERE col > $last ORDER BY col ASC), persisting the last read value under cursor_id. Captures appends only — updates and deletes are not observed. Available on SQLx (PostgreSQL / MySQL / MariaDB / SQLite) and ClickHouse; while idle the poll interval backs off exponentially between polling_interval_ms and max_polling_interval_ms. SQLite and ClickHouse are polling-only — they have no server-side change log.

SQLx read mode is chosen by config, not by driver. With no cursor_column, an SQLx source (?table= / table:) is a competing-consumers work queue: it atomically claims rows via a locked_until lease and deletes them on ack, so several readers drain the same table as competing consumers — each row goes to one reader at a time, at-least-once (a lease that expires before ack, e.g. after a crash, redelivers the row), not exactly once. This requires the table to carry the queue schema — an id, a payload, and a locked_until column — on every driver (PostgreSQL, MySQL/MariaDB, and SQLite behave identically here); it is not a plain full-table read. To read an existing or arbitrary table non-destructively (the ETL path), set cursor_column to switch to the cursor polling above. A source table missing locked_until fails fast with a permanent error rather than reconnecting forever.

Choosing a MongoDB consume mode

These are not interchangeable. The default capture_all is a reader, like the other capture_* modes: they fan out, so every reader sees every document, and nothing is claimed or removed. Opting into consumer instead gives a competing-consumers work queue: its lock-and-delete protocol lets several readers drain the same collection with each document going to exactly one of them — the right thing for dispatching jobs or commands, and destructive by design.

Pick by semantics first. Where both would do — a one-shot bulk read or ETL pass with a single reader — use capture_all, the default: consumer's four round trips per batch (find ids → claim → re-fetch → delete on ack) buy exclusivity you aren't using. Set consume: consumer explicitly when you want the destructive work-queue semantics.

500k documents, batch_size: 1024, release build, standalone MongoDB: consumer → null 23,667 rows/s (postgres → null reference: 115,924). The previously quoted capture_all figure was measured on that same standalone server, i.e. through the _id fallback that has since been removed as unsound — it is not comparable and needs re-measuring on a replica set.

• concurrency does not speed up a MongoDB source — batches are fetched serially; it only widens the downstream side.
• consumer only reads collections written by the mq-bridge MongoDB publisher (UUID _id / wrapped envelope); other documents are skipped or rejected at startup. snapshot and the capture_* modes read arbitrary collections.
• capture_* need a replica set — a single-node one is enough — and both error without it. With exit_on_empty / --drain, a replica-set reader finishes after the snapshot and any immediately following changes, once the drain idle timeout expires.
• snapshot is one-shot by design: it delivers what exists when the run starts and then ends the route. It rejects cursor_id, because resuming above a stored _id would skip anything a concurrent writer commits below that mark — see the note below.
• Wart: capture_all's snapshot also emits the internal <collection>:sequencer document.

Why snapshot does not resume. A _id cursor can only match documents above its high-water mark, and _id is assigned client-side before the insert, so it does not follow commit order. Any writer that commits below the mark — concurrent producers, a route with concurrency > 1, a retried write — is skipped permanently and silently. Incremental reads therefore need commit order, which on MongoDB means the oplog, which means a replica set. There is no sound _id-based substitute.

Deduplication & idempotent writes

With a durable source and durable checkpoint configuration, mq-bridge is at-least-once across crashes: a replay can redeliver, while in-process Memory endpoint state is not crash-durable. Pair a stable replay identity with an idempotent sink operation and that covered sink effect is effectively-once — exactly-once effects, however often the message is delivered. Which sink absorbs a duplicate, which source gives you a stable key to deduplicate on, and what a handler in the route changes about all of this, is covered in full by:

docs/DELIVERY.md — delivery guarantees. Per-source identity and per-sink idempotency tables, the deduplication middleware, Mongo id_field and report_outcome, SQL ON CONFLICT, ClickHouse ReplacingMergeTree, file/object-store name_by: source_position, what handlers change, and what mq-bridge deliberately does not provide.

The short form: point the sink at a deterministic business key (id_field for MongoDB, an ON CONFLICT column for SQL) and let its unique constraint be the authority — it is already shared across every writer, so no extra state store is needed. Add the deduplication middleware on the input only when the sink has no constraint to lean on, or to keep duplicates away from a handler.

Cloud Object Storage (S3 / GCS / Azure)

The object_store endpoint (alias s3) reads and writes cloud object stores — Amazon S3, Google Cloud Storage, Azure Blob, Cloudflare R2, and anything else the object_store crate speaks — behind the same receive_batch / send_batch API. Enable it with the object-store feature. Credentials and backend options are read from the process environment (AWS_ACCESS_KEY_ID, AWS_ENDPOINT, AWS_REGION, GOOGLE_SERVICE_ACCOUNT, AZURE_STORAGE_ACCOUNT, ...); the URL scheme picks the backend (s3://, gs://, az://).

• As a sink, batches are encoded with the same file endpoint formats (normal JSONL, json, text, raw) and written as whole immutable objects — write-once, nothing appended or mutated. Under name_by: write_time each flushed batch becomes one object at <prefix>/[YYYY/MM/DD/]<uuidv7>.<ext>; the uuidv7 name already sorts by write time, and the optional date_partition prefix (on by default, derived from that same id's timestamp) is a readability / lifecycle-rule convenience; date_partition_style: hive writes it as year=YYYY/month=MM/day=DD/, which Spark, DuckDB, DataFusion and Athena read as partition columns. Under name_by: source_position a batch becomes one object per contiguous run of source positions it covers — usually one, more when the batch spans partitions or skips offsets a replay already wrote — each named for the range it covers and written flat under the prefix, where date_partition does not apply. The default name_by: auto picks source_position whenever the route's input carries a replay position — Kafka, Postgres CDC, a SQL cursor, a MongoDB change stream or a file — so key order equals source order at any concurrency without configuring anything (see Files & object storage — name_by).
• As a source, objects under the prefix are listed in key order, fetched whole, split on the delimiter, and emitted. Progress is a durable cursor holding the last fully-acked object key: set cursor_id and an external checkpoint_store (file://, s3://, postgres://, mongodb://) so a restart resumes without re-emitting. Objects are never deleted or rewritten — resume is non-destructive and at-least-once at object granularity (a nacked batch is redelivered; the cursor only advances once an object is fully acked). csv is supported on the source only.
• Parquet (format: parquet, parquet feature, included in full): each batch becomes one .parquet object whose schema is inferred from that batch, so DataFusion, DuckDB and other lake engines can query the prefix directly. Every payload must be a JSON object; any other payload fails on its own — sent to the DLQ when the route has dlq middleware, otherwise acknowledged and dropped — and the rest of the batch is still written. On a Parquet object, compression picks the column codec (gzip, lz4 = LZ4_RAW, zstd) instead of wrapping the whole file, so the extension stays .parquet. As a source it emits one JSON-object message per row, with typed values, and it also reads snappy-compressed files written by other tools. Because each object's column types come from its own batch, put a transform with a schema before the sink when a prefix is queried as one table: it pins every field's type (a value is coerced losslessly or the message is rejected), so a value that is a number in one batch and a string in the next can't give the objects conflicting schemas. A field that is null in every row of a batch is still written as an untyped (NULL) column; DataFusion merges that with the other objects, and DuckDB needs read_parquet(..., union_by_name = true). The file endpoint rejects parquet because a Parquet file can't be appended to.

archive_to_s3:
  input:
    memory: { topic: "events" }
  output:
    object_store:
      url: "s3://my-bucket/events"
      format: normal            # JSONL; also json / text / raw

replay_from_s3:
  input:
    object_store:
      url: "s3://my-bucket/events"
      cursor_id: "replayer-1"
      checkpoint_store: "file:///var/lib/mqb/s3-cursor.json"
  output:
    nats: { subject: "events.replay", url: "nats://localhost:4222" }

Point checkpoint_store at a different bucket or prefix than the source reads; a cursor object written under the source prefix would be listed and re-read as data (the source rejects an overlapping object-store checkpoint location).

The sink's format decides the encoding, not where the message came from: normal/json/text always write the {message_id, payload, metadata} wrapper, so the message id survives the round trip. Use format: raw to write payloads verbatim (bare documents, no wrapper). Applies to both file and object_store.

A source on those formats expects that same wrapper, message_id included. A line that is valid JSON but not the wrapper — a hand-written fixture with only payload and metadata, say — is not decomposed: the whole line becomes the payload and the line's own metadata is discarded. The reader logs a warning naming the wrapper when this happens. For plain JSON lines, use format: raw.

Payload encoding in the wrapper. A UTF-8 payload is written as a plain JSON string under payload. A binary one (compressed, encrypted, Protobuf, …) is base64-encoded under a separate payload_base64 field; the two are mutually exclusive, as in the CloudEvents JSON format.

json

{"message_id":"019f9b12-d786-7ebe-a7ec-a1aa71bc47ae","payload":"{\"order_id\":7}"}
{"message_id":"019f9b12-d78a-7c01-b0f4-2f0f4d6a1c33","payload_base64":"KLUv/SBOAQAA"}

Sources still read the older byte-array form ("payload":[123,34,…]), so existing files keep working. Only the reverse is a break: a binary record written by this version is not readable by an older mq-bridge. Text records are compatible in both directions.

Response Endpoint

The response output endpoint sends a reply back to the original requester. This is useful for synchronous request-reply flows, for example HTTP-to-NATS-to-HTTP. Use response: {} as the output endpoint configuration.

• Caveats:
  • If the input does not support responses (e.g., File, SQLx), the message sent to response will be dropped.
  • Ensure timeouts are configured correctly on the requester side, as the bridge processing time adds latency.
  • Middleware that drops metadata (like correlation_id) may break the response chain.

Usage

The mq-bridge-app workspace packages run mq-bridge as a standalone CLI, desktop app, or Docker container configured via YAML or environment variables. Plain cargo build still builds only this library; use cargo build --workspace to build every workspace package.

Configuration-first workflow

mq-bridge-app can be used to create and test route and endpoint configurations through a UI. The generated JSON/YAML can then be copied into an application and loaded by the Rust library or the Python bindings.

This does not replace application code or handlers, but it is useful when you want a known-good connection and route shape before pasting the configuration into code. For Python and Node projects, routes and publishers are commonly loaded from JSON/YAML with Route.from_file/fromFile (a file path) and Route.from_str/fromStr (a JSON/YAML string), while Route.from_config/fromConfig takes an already-parsed dict/JS object. The matching Publisher.* constructors work the same way.

Programmatic Handlers

For business logic, mq-bridge provides a handler layer separate from transport-level middleware. This is where message-specific code usually belongs.

Raw Handlers

• CommandHandler: A handler for 1-to-1 or 1-to-0 message transformations. It takes a message and can optionally return a new message to be passed down the publisher chain.
• EventHandler: A terminal handler that reads new messages without removing them for other event handlers.

You can chain these handlers with endpoint publishers.

use mq_bridge::traits::Handler;
use mq_bridge::{CanonicalMessage, Handled};
use std::sync::Arc;

// Define a handler that transforms the message payload
let command_handler = |mut msg: CanonicalMessage| async move {
    let new_payload = format!("handled_{}", String::from_utf8_lossy(&msg.payload));
    msg.payload = new_payload.into();
    Ok(Handled::Publish(msg))
};

// Attach the handler to a route
// let route = Route { ... }.with_handler(command_handler);

Typed Handlers

For more structured, type-safe message handling, mq-bridge provides TypeHandler. It deserializes messages into a specific Rust type before passing them to a handler function, so handlers do not need to repeat the same parsing code.

Message selection is based on the kind metadata field in the CanonicalMessage.

use mq_bridge::type_handler::TypeHandler;
use mq_bridge::{CanonicalMessage, Handled};
use serde::Deserialize;
use std::sync::Arc;

// 1. Define your message structures
#[derive(Deserialize)]
struct CreateUser {
    id: u32,
    username: String,
}

#[derive(Deserialize)]
struct DeleteUser {
    id: u32,
}

// 2. Create a TypeHandler and register your typed handlers
let typed_handler = TypeHandler::new()
    .add("create_user", |cmd: CreateUser| async move {
        println!("Handling create_user: {}, {}", cmd.id, cmd.username);
        // Logic here...
        // Automatically maps () to Handled::Ack
    })
    .add("delete_user", |cmd: DeleteUser| async move {
        println!("Handling delete_user: {}", cmd.id);
        // Logic here...
        // Automatically maps () to Handled::Ack
    });

// 3. Attach the handler to a route
let route = Route::new(input, output).with_handler(typed_handler);

// 4. To send a message to the route's input, create a publisher for that endpoint.
//    In a real application, you would create this publisher once and reuse it.
let input_publisher = Publisher::new(route.input.clone()).await.unwrap();

// 5. Create a typed command, serialize it, and send it via the publisher.
let command = CreateUser { id: 1, username: "test".to_string() };
let message = msg!(&command, "create_user"); // This sets the `kind` metadata field.
input_publisher.send(message).await.expect("Failed to send message");

// The running route will receive the message, see the `kind: "create_user"` metadata,
// deserialize the payload into a `CreateUser` struct, and pass it to your registered handler.

Programmatic Usage

You can define and run routes directly in Rust code.

use mq_bridge::{models::Endpoint, stop_route, CanonicalMessage, Handled, Route};
use std::sync::{
    atomic::{AtomicBool, Ordering},
    Arc,
};

#[tokio::main]
async fn main() {
    // Define a route from one in-memory channel to another

    // 1. Create a boolean that is changed in the handler
    let success = Arc::new(AtomicBool::new(false));
    let success_clone = success.clone();

    // 2. Define the Handler
    let handler = move |mut msg: CanonicalMessage| {
        success_clone.store(true, Ordering::SeqCst);
        msg.set_payload_str(format!("modified {}", msg.get_payload_str()));
        async move { Ok(Handled::Publish(msg)) }
    };
    // 3. Define Route
    let input = Endpoint::new_memory("route_in", 200);
    let output = Endpoint::new_memory("route_out", 200);
    let route = Route::new(input, output).with_handler(handler);

    // 4. Run (deploys the route in the background)
    route.deploy("test_route").await.unwrap();

    // 5. Inject Data
    let input_channel = route.input.channel().unwrap();
    input_channel
        .send_message("hello".into())
        .await
        .unwrap();

    // 6. Verify
    let mut verifier = route.connect_to_output("verifier").await.unwrap();
    let received = verifier.receive().await.unwrap();
    assert_eq!(received.message.get_payload_str(), "modified hello");
    assert!(success.load(Ordering::SeqCst));

    stop_route("test_route").await;
}

Patterns: Request-Response

mq-bridge supports request-response patterns for interactive services such as web APIs. A client can send a request and wait for the matching response, while the bridge keeps the correlation details away from the handler.

The response output is the most direct option and the safest one under concurrency.

The response Output Endpoint (Recommended)

For request-response routes, use the dedicated response endpoint in the route's output.

How it works:

1. An input endpoint that supports request-response (like http) receives a request.
2. The message is passed through the route's processing chain. This is where you typically attach a handler to process the request and generate a response payload.
3. The final message is sent to the output.
4. If the output is response: {}, the bridge sends the message back to the original input source, which then sends it as the reply (e.g., as an HTTP response).

The response stays in the same execution context as the request, so concurrent requests do not need to share a reply queue and race on correlation IDs.

Example: MongoDB Request-Response

For example, a service can write a request document to MongoDB and wait for a reply. The bridge reads the document, runs the handler, and writes the result back to the reply collection.

YAML Configuration (mq-bridge.yaml):

mongo_responder:
  input:
    mongodb:
      url: "mongodb://localhost:27017"
      database: "app_db"
      collection: "requests"
  output:
    # The 'response' endpoint sends the processed message back to the 'requests_replies' collection
    # (or whatever reply_to was set to by the sender).
    response: {}

Programmatic Handler Attachment (in Rust): You would then load this configuration and attach a handler to the route's output endpoint in your Rust code.

use mq_bridge::models::{Config, Handled};
use mq_bridge::CanonicalMessage;

async fn run() {
    // 1. Load configuration from YAML
    // let config: Config = serde_yaml_ng::from_str(include_str!("mq-bridge.yaml")).unwrap();
    // let mut route = config.get("api_gateway").unwrap().clone();

    // 2. Define the handler that processes the request
    let handler = |mut msg: CanonicalMessage| async move {
        // Example: echo the request body with a prefix
        let request_body = String::from_utf8_lossy(&msg.payload);
        let response_body = format!("Handled response for: {}", request_body);
        msg.payload = response_body.into();
        Ok(Handled::Publish(msg))
    };

    // 3. Attach the handler to the output endpoint
    // route.output.handler = Some(std::sync::Arc::new(handler));

    // 4. Run the route
    // route.deploy("api_gateway").await.unwrap();
}

Patterns: CQRS

mq-bridge can be used for CQRS-style flows. With routes and typed handlers, it can act as a command bus and an event bus without becoming a domain framework.

• Command Bus: An input source (e.g., HTTP) receives a command. A TypeHandler processes it (Write Model) and optionally emits an event.
• Event Bus: The emitted event is published to a broker (e.g., Kafka). Downstream routes subscribe to these events to update Read Models (Projections).

// 1. Command Handler (Write Side)
let command_bus = TypeHandler::new()
    .add("submit_order", |cmd: SubmitOrder| async move {
        // Execute business logic, save to DB...
        // Emit event
        let evt = OrderSubmitted { id: cmd.id };
        Ok(Handled::Publish(
            msg!(evt, "order_submitted")
        ))
});

// 2. Event Handler (Read Side / Projection)
let projection_handler = TypeHandler::new()
    .add("order_submitted", |evt: OrderSubmitted| async move {
        // Update read database / cache...
        // Ok(()) is equivalent to Handled::Ack
        Ok(())
});

Configuration

All routes and endpoints can be defined via a configuration file (for example mq-bridge.yaml), JSON, or environment variables. For a complete reference of options, middleware, and examples, see the Configuration Guide.

Important route-level knobs:

• batch_size: maximum messages per route iteration. Defaults to 512; lower it for tighter failure isolation or large payloads.
• concurrency: number of route workers. Defaults to 1; useful for high-latency handlers or endpoints.
• commit_concurrency_limit: maximum in-flight commit operations, whether they are queued through ordered commit sequencing or run concurrently for independent-ack transports. Defaults to 4096.

Middleware can be attached to inputs or outputs. The most commonly used ones are retry, dlq, deduplication, limiter, cookie_jar, and transform. Retry/DLQ are especially useful with batching because partial failures can be retried or sent to a DLQ without treating the entire batch as equally broken.

Transforming JSON messages

The transform middleware reshapes JSON payloads declaratively, so field renaming and type fixing do not need a custom handler. It runs two optional stages over a single parse: mapping (rename, move, nest) and then schema (coerce, apply defaults, validate).

csv_to_kafka:
  input:
    file: { path: "users.csv", format: csv }
  output:
    middlewares:
      - transform:
          mapping:
            firstName: "$.first_name"
            lastName: "$.last_name"
            id: "$.user_id"
            "address.city": { path: "$.city", default: "unknown" }
          schema_file: "schemas/user.json"
      # `dlq` comes after `transform`: publisher middlewares are wrapped in list order,
      # so the last entry is the outermost layer and sees the earlier ones' failures.
      - dlq:
          endpoint: { file: { path: "rejected.jsonl" } }
    kafka: { topic: "users", url: "localhost:9092" }

With the CSV source above producing {"first_name":"John","last_name":"Smith","user_id":"42"}, the mapping yields {"firstName":"John","lastName":"Smith","id":"42","address":{"city":"unknown"}} ($.city is absent, so address.city falls back to its "unknown" default) and the schema then coerces id to the integer 42.

Failures are always non-retryable and name the offending field, e.g. transform failed at $.items[1].qty [coercion]: cannot coerce string "oops" to integer. On an output endpoint the message is failed so a following dlq captures it; on an input endpoint it is dropped from the batch and acknowledged, which keeps invalid data out of the route.

See REFERENCE.md for the full option list, the supported schema subset, and the exact set of allowed type coercions.

Running Tests

The project includes integration and performance tests. Most backend tests require Docker.

To run the performance benchmarks for all supported backends:

cargo test --test integration_test --release -- --ignored --nocapture --test-threads=1

To run the criterion benchmarks:

cargo bench --features "full"

The times are not stable yet, it is therefore recommended to perform the integration performance test if you want to measure throughput.

Contributing

Contributions are welcome. See CONTRIBUTING.md for setup notes, code style, and pull request guidelines.

AI Disclaimer

This library has been written with a lot of AI assistance.

The core started as my own code, and many endpoints and docs were expanded with help from Gemini, CodeRabbit, Claude, and Codex. The useful part was speed: once the endpoint traits were stable, adding more transports became much easier. The dangerous part is the usual one: generated code can look plausible while still missing important details. I am aware that in year 2026, AI is still not generating perfect code and sometimes breaks simple stuff or forgets important lines during refactorings that later result in severe bugs.

For that reason I reviewed each commit manually to prevent hard-to-fix architectural issuess and cleaned up, and refactored the generated output.

I do trust the current code as much as if it would be written by myself.

Due to the large feature set, there may still be unfixed issues. The current focus is testing and documentation.

Some parts of the code are more verbose than I would write by hand, but I kept the readable parts when they worked well. I am not a native English speaker, so AI assistance is also useful for documentation. The important part is that the code is reviewed and tested, not that every sentence or helper function looks hand-typed from the first draft.

License

mq-bridge is dual-licensed under either the MIT license or the Apache License, Version 2.0, at your option.

Binary distributions carry their applicable third-party terms in THIRD_PARTY_LICENSES.txt. Their current maintenance process is documented in docs/THIRD_PARTY_LICENSES.md.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in mq-bridge by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.