Skip to content

0.3.0

Compare
Choose a tag to compare
@aphyr aphyr released this 06 Dec 04:48
· 85 commits to main since this release

This release replaces many of Jepsen's internals with faster or more scalable data structures. It introduces significant new datatypes and adds new support libraries. Core generators are much faster, thanks to new Context and Op types. Running and analyzing tests can be 1-2 orders of magnitude faster: Jepsen can now run list-append tests at ~45,000 ops/sec and check them at ~30,000 ops/sec. Histories are streamed and loaded incrementally, which improves crash recovery, allows for histories larger than RAM, and speeds up REPL work. Histories in the hundreds of millions or even billions of operations are now tractable. Most checkers are parallelized and take advantage of sophisticated multi-query optimization for reductions over histories. A new dependency-aware executor allows checkers to run in parallel without starvation. New nemesis.combined packages support file truncation and bitflips, as well as network latency and packet loss.

As usual, most things should be API compatible, and we try to issue Obvious Warnings when they're not--but this is a big enough change that we're bumping the minor version from 0.2.7 to 0.3.0. Users integrating tightly with histories and generators should test their code carefully.

New Features

  • A new library, jepsen.history, provides support for writing efficient checkers. It includes a transactional dependency-aware concurrent executor, concurrent and linear folds with multi-query optimization, and lazy datatypes for working with large histories.
  • Operations are now represented by an Op defrecord (jepsen.history.Op) instead of maps. This yields significant performance and speed improvements. Ops have mandatory :index and :time fields, both longs. See jepsen.history for more details.
  • Histories are incrementally streamed to the test.jepsen file, and sealed in 16384-operation chunks. If a test crashes during the run or analysis phase, you can likely recover some of its history and re-analyze it.
  • Histories are now represented by subtypes of jepsen.history.History. These should be compatible with vectors, but stream their contents lazily from disk. Mapping between invocations and completions is now built in to histories, rather than being an external pair-index structure. Histories support efficient linear and concurrent folds with stream fusion and multi-query optimization, and directly support Tesser folds. Analyses may be 1-2 orders of magnitude faster, depending on hardware. See jepsen.history for details.
  • dom-top.core has a new reducer macro which roughly doubles performance for reductions with multiple accumulator variables.
  • Elle can catch new classes of anomalies, especially involving realtime and process-including anti-dependency cycles.
  • lein run analyze now pulls the test arguments out of the test; you don't have to pass them every time.
  • A new nemesis.combined/file-corruption-package provides support for bitflips and truncation of files.
  • A new nemesis.combined/packet-package induces network latency and packet loss.
  • A new tests.kafka namespace supports tests for Kafka-style append-only ordered logs.
  • util/rand-distribution supports picking random numbers

Significant API Changes

  • Operations are now jepsen.history.Ops, not maps. :index and :time fields are now mandatory.
  • Histories are now subtypes of jepsen.history.History, not vectors. They should be mostly API compatible, and will transparently promote themselves to vectors on certain operations (for instance, conj).
  • Generator contexts are now jepsen.context.Contexts, rather than maps. Accessing their old fields will throw and warn you to use new polymorphic functions in jepsen.context.
  • lein run analyze now takes -t path-to-test or -t test-index, rather than the full arguments to recreate the test map.
  • test.fressian files, deprecated in 0.2.x, are no longer generated. Use test.jepsen instead.

Performance Improvements

  • Accessing operations is much faster thanks to jepsen.history.Op
  • jepsen.generator is roughly an order of magnitude faster, especially for high (~thousands of threads) concurrency tests, thanks to the new generator.context.Context type.
  • Generators can now dynamically compile context-filtering operations to BitSet intersections, which speeds up reserve, on-threads, clients, nemesis, and other generators.
  • Reductions over histories (e.g. basically every checker) are 1-2 orders of magnitude faster, thanks to jepsen.history.
  • Elle is roughly an order of magnitude faster, thanks to jepsen.history and careful parallelization.
  • Assorted optimizations to generator/fill-in-op, soonest-op-mop, and reserve make them significantly faster.
  • Tests no longer need to wait for history writing at the end of the test, since it's streamed to disk.
  • Using functions as generators is now faster; we perform arity reflection only once rather than on every op.
  • store.fressian decodes lists as vectors directly, rather than post-processing them. This makes Fressian decoding significantly faster.

Minor Improvements

  • Jepsen and Elle used knossos.history and knossos.op extensively. These have been almost entirely replaced with jepsen.history.
  • Most checkers have been rewritten to use jepsen.history; many reductions are now concurrent folds.
  • Knossos 0.3.9
  • Tools.cli 1.0.214
  • Unilog 0.7.31
  • Ring 1.9.6
  • SSHJ 0.34.0
  • Elle 0.1.6
  • Lazyfs c16518f6
  • Assorted type hints and compiler warnings resolved
  • Contexts are deterministic again, rather than stochastic. This may break tests that depended on specific nondeterministic orders.

Full Changelog: v0.2.7...v0.3.0