Jepsen tests for the etcd consensus system! To run these, you'll need a Jepsen cluster.

To run a full test suite, run

lein run test-all --concurrency 2n --rate 1000

You can focus in on particular combinations of faults, or particular workloads:

lein run test-all --concurrency 2n --workload append
lein run test-all --concurrency 2n --nemesis kill,partition

You can also run particular workloads and nemeses with lein run test. For instance, to demonstrate lost updates to a set protected by an etcd lock, try:

lein run test --workload lock-set --nemesis pause --time-limit 120

append appends values to keys and reads them back, using transactions to perform multiple reads/appends atomically. Each append is implemented by reading the current value of the key from etcd, and writing it back using a compare guard on the observed revision, to ensure nobody has changed it during the transaction.

lock-set uses an etcd lock to protect access to a shared in-memory set stored on the local control node, inside Jepsen itself. lock-etcd-set does the same, but stores the set in etcd. Both of these are unsafe, because etcd locks are fundamentally unsafe.

set appends unique elements to a single key using compare-and-set transactions, and tries to read the set back.

register tests for linearizable reads, writes, and compare-and-set operations on named registers. Each register is checked independently, since there are no cross-register transactions. The checker is aware of revision information as well as the actual value of the register.

watch has processes observe change events on a single key as it goes through various revisions, then confirms that everyone observed the same state transitions in the same order. Watch has a tendency to get stuck if we clobber the cluster too badly, which may result in indefinite results.

pause, kill, clock, and partition are the usual Jepsen nemeses for pausing/killing etcd processes, adjusting and strobing the clock, and partitioning the network. member adds and removes members dynamically. bitflip-wal and bitflip-snap induce single-bit errors in the WAL and snapshot files on disk.

--concurrency 4n says "run 4 worker threads per node".

--lazyfs mounts the etcd data directory in a lazyfs FUSE filesystem, and alters the kill nemesis to also lose un-fsynced writes.

--ops-per-key controls how many operations we perform on a single key.

--nemesis FAULTS takes a comma-separated list of faults to inject. --nemesis-interval SECONDS sets the approximate time between nemesis operations.

--rate HZ sets the rough number of operations per second Jepsen tries to perform.

--serializable allows etcd to serve stale data.

--snapshot-count N controls the snapshot count parameter to etcd; we lower this to 100 by default to stress the snapshot mechanism.

--tcpdump captures client traffic to each node in a pcap file.

--unsafe-no-fsync tells etcd to not perform fsyncs. This should be unsafe.

--time-limit SECONDS controls how long to run the test for.

Copyright © 2019, 2020, 2022 Jepsen, LLC

This program and the accompanying materials are made available under the terms of the Eclipse Public License 2.0 which is available at http://www.eclipse.org/legal/epl-2.0.

This Source Code may also be made available under the following Secondary Licenses when the conditions for such availability set forth in the Eclipse Public License, v. 2.0 are satisfied: GNU General Public License as published by the Free Software Foundation, either version 2 of the License, or (at your option) any later version, with the GNU Classpath Exception which is available at https://www.gnu.org/software/classpath/license.html.