Maelstrom is a workbench for learning distributed systems by writing your own. It uses the Jepsen testing library to test toy implementations of distributed systems. Maelstrom provides standardized tests for things like "a commutative set" or "a transactional key-value store", and lets you learn by writing implementations which those test suites can exercise. It's used as a part of a distributed systems workshop by Jepsen.
Maelstrom provides a range of tests for different kinds of distributed systems, built on top of a simple JSON protocol via STDIN and STDOUT. Users write servers in any language. Maelstrom runs those servers, sends them requests, routes messages via a simulated network, and checks that clients observe expected behavior. You want to write Plumtree in Bash? Byzantine Paxos in Intercal? Maelstrom is for you.
Maelstrom's tooling lets users experiment with simulated latency and message loss. Every test includes timeline visualizations of concurrency structure, statistics on messages exchanged, timeseries graphs to understand how latency, availability, and throughput respond to changing conditions, and Lamport diagrams so you can understand exactly how messages flow through your system. Maelstrom's checkers can verify sophisticated safety properties up to strict serializability, and generate intuitive, minimal examples of consistency anomalies.
Maelstrom can help you model how a system responds to different cluster sizes, network topologies, latency distributions, and faults like network partitions. Maelstrom also offers simulated services that you can use to build more complex systems.
It's built for testing toy systems, but don't let that fool you: Maelstrom is reasonably fast and can handle simulated clusters of 25+ nodes. On a 48-way Xeon, it can use 94% of all cores, pushing upwards of 60,000 network messages/sec.
The Maelstrom Guide will take you through writing several different types of distributed algorithms using Maelstrom. We begin by setting up Maelstrom and its dependencies, write our own tiny echo server, and move on to more sophisticated workloads.
- Chapter 1: Getting Ready
- Chapter 2: Echo
- Chapter 3: Broadcast
- Chapter 4: CRDTs
- Chapter 5: Datomic
- Chapter 6: Raft
There are also several reference documents that may be helpful:
- Protocol defines Maelstrom's "network" protocol, message structure, RPC semantics, and error handling.
- Workloads describes the various kinds of workloads that Maelstrom can test, and define the messages involved in that particular workload.
- Understanding Test Results explains how to interpret the various plots, data structures, and log files generated by a test.
- Services discusses Maelstrom's built-in network services, which you can use as primitives in building more complex systems.
Writing "real" distributed systems involves a lot of busywork: process management, networking, and message serialization are complex, full of edge cases, and difficult to debug across languages. In addition, running a full cluster of virtual machines connected by a real IP network is tricky for many users. Maelstrom strips these problems away so you can focus on the algorithmic essentials: process state, transitions, and messages.
The "nodes" in a Maelstrom test are plain old binaries written in any language. Nodes read "network" messages as JSON from STDIN, write JSON "network" messages to STDOUT, and do their logging to STDERR. Maelstrom runs those nodes as processes on your local machine, and connects them via a simulated network. Maelstrom runs a collection of simulated network clients which make requests to those nodes, receive responses, and records a history of those operations. At the end of a test run, Maelstrom analyzes that history to identify safety violations.
This allows learners to write their nodes in whatever language they are most comfortable with, without having to worry about discovery, network communication, daemonization, writing their own distributed test harness, and so on. It also means that Maelstrom can perform sophisticated fault injection and trace analysis.
Maelstrom starts in
maelstrom.core, which parses CLI operations and
constructs a test map. It hands that off to Jepsen, which sets up the servers
and Maelstrom services via
maelstrom.db. Spawning binaries and handling their
IO is done in
maelstrom.process, and Maelstrom's internal services (e.g.
lin-kv are defined in
maelstrom.service. Jepsen then spawns clients
depending on the workload (
maelstrom.workload.*) and a nemesis
maelstrom.nemesis) to inject faults.
Messages between nodes are routed by
maelstrom.net, and logged in
maelstrom.journal. Clients send requests and parse responses via
maelstrom.client, which also defines Maelstrom's RPC protocol.
At the end of the test, Jepsen checks the history using a checker built in
maelstrom.core: workload-specific checkers are defined in their respective
workload namespaces. Network statistics are computed in
maelstrom.net.journal, and Lamport diagrams are generated by
maelstrom.doc helps generate documentation based on
registry of RPC types and workloads.
A full list of options is available by running
java -jar maelstrom.jar test --help. The important ones are:
--workload NAME: What kind of workload should be run?
--bin SOME_BINARY: The program you'd like Maelstrom to spawn instances of
--node-count NODE-NAME: How many instances of the binary should be spawned?
To get more information, use:
--log-stderr: Show STDERR output from each node in the Maelstrom log
--log-net-send: Log messages as they are sent into the network
--log-net-recv: Log messages as they are received by nodes
To make tests more or less aggressive, use:
--time-limit SECONDS: How long to run tests for
--rate FLOAT: Approximate number of requests per second
--concurrency INT: Number of clients to run concurrently. Use
4nfor 4 times the number of nodes.
--latency MILLIS: Approximate simulated network latency, during normal operations.
--latency-dist DIST: What latency distribution should Maelstrom use?
--nemesis FAULT_TYPE: A comma-separated list of faults to inject
--nemesis-interval SECONDS: How long between nemesis operations, on average
For broadcast tests, try
--topology TYPE: Controls the shape of the network topology Jepsen offers to nodes
For transactional tests, you can control transaction generation using
--max-txn-length INT: The maximum number of operations per transaction
--key-count INT: The number of concurrent keys to work with
--max-writes-per-key INT: How many unique write operations to generate per key.
SSH options are unused; Maelstrom runs entirely on the local node.
Raft node processes still alive after maelstrom run
You may find that node processes maelstrom starts are not terminating at the end of a run as expected. To address this, make sure that if the process passed as
--bin forks off a new process, it also handles the process' termination.
#!/bin/bash # Forks a new process. java -jar target/raft.jar
#!/bin/bash # Replaces the shell without creating a new process. exec java -jar target/raft.jar
Copyright © 2017, 2020, 2021 Kyle Kingsbury, Kit Patella, & Jepsen, LLC
Distributed under the Eclipse Public License either version 1.0 or (at your option) any later version.