Jepsen tests distributed and concurrent systems safety. To analyze these systems, Jepsen uses a history: a totally ordered log of concurrent logical operations. This Clojure library provides support for various tools (including Jepsen itself) that work with histories. Its documentation also describes the structure and semantics of histories, so that other people can interpret them.
- jepsen.history provides an Op datatype and a family of history types with efficient get-by-index, invocation-completion mapping, optimized concurrent folds, and memory-efficient, lazy map/filter, plus dependency-aware futures. It uses...
- jepsen.history.fold is a stateful executor for running linear and concurrent folds over large chunked collections (like histories). It automatically fuses together concurrent folds into fewer passes over the underlying dataset.
- jepsen.history.task offers a dependency-aware, transactional ThreadPoolExecutor for compute-bound tasks (like analysis of histories).
A short, real history from a transactional test of a system might be:
[{:type :invoke, :f :txn, :value [[:w 2 1]], :time 3291485317, :process 0, :index 0}
{:type :invoke, :f :txn, :value [[:r 0 nil] [:w 1 1] [:r 2 nil] [:w 1 2]], :time 3296209422, :process 2, :index 1}
{:type :fail, :f :txn, :value [[:r 0 nil] [:w 1 1] [:r 2 nil] [:w 1 2]], :time 3565403674, :process 2, :index 2, :error [:duplicate-key "etcdserver: duplicate key given in txn request"]}
{:type :ok, :f :txn, :value [[:w 2 1]], :time 3767733708, :process 0, :index 3}]This shows two concurrent transactions executed by process 0 and 2, respectively; process 0's transaction attempted to write key 2 = 1 and succeeded, but process 2's transaction tried to read 0, write key 1 = 1, read 2, then write key 1 = 2, and failed.
Each element of a history is called an operation. Operations are maps with several well-defined keys:
The :type must be one of :invoke, :ok, :info, or :fail. :invoke
denotes the start of a logical operation; :ok denotes its successful
completion. :fail denotes its definite failure. :info denotes an
indeterminate result: the operation either may or may not have completed. For
historical reasons, :info is also used for all operations performed by the
:nemesis process, which performs fault injection. This has been somewhat
awkward, and may change in the future.
The :time is a long: The time since the start of the test, in nanoseconds.
The :process is a logical identifier for the process executing this
operation, like 2 or :nemesis. Processes are logically singlethreaded, and
can execute only one thing at a time. If a process crashes (e.g. by beginning
an :invoke followed by an :info) it can never invoke an operation again.
The :f is a function being applied. This may be any object, but is often
something like :read, :write, :txn, :dequeue, :kill, etc. The
interpretation of :f is generally up to the user.
The :value stores arguments to and/or return values from the operation. For
instance, the value of a read might be the value read from the database (and
for the invocation of the read, is typically left nil. For a write, it might
be the value written to the database.
The :index is a unique, monotonically ascending integer identifying this
operation in the history. In most histories these are usually 0, 1, 2, ...,
which is helpful for encoding operations in an array or integer-based data
structure, then mapping back to full operations later. However, users may omit
indices (e.g. for testing), or filter a history to a subset of operations
(which makes indices sparse). We make efforts to support both cases.
An :error field encodes error information, and is often helpful for debugging
why an operation returned :info or :fail.
Operations may contain any number of additional fields, which users are free to
interpret as they like. For instance, one might store :debug information
related to the operation's execution, or the :version of the database
interacted with.
There are two basic flavors of process: clients, which have integer
:process identifiers, and nemeses, which have other :process
identifiers--for instance :process :nemesis. Clients interact with the
logical state of the system. Nemeses introduce faults and other meta-level
operations into a system, and are not supposed to have any impact on logical
safety.
An operation which completes with :ok must logically affect the system. An operation which completes with :fail must have no logical effect on the system. An :info may do either.
For historical reasons, nemeses use :info when they invoke, and :info when
they complete an operation. This is sort of awkward, but changing it will
require reviewing huge swaths of checker code that assumes :infos can't
affect system state.
The logical semantics of an operation--what it would mean to execute that
operation in an abstract model of the system--should, in general, depend only
on its :f and :value.
Checkers often need to do things like:
-
Compute a derived history based on an input history, perform an analysis on that derived history, then map the operations we find back to operations in the original history for presentation.
-
Reindex a history which has sparse or missing indices
-
Access an operation in a history by index, even if indices are sparse
-
Map each operation of a history into a new kind of operation
-
Filter a history to just operations from clients, or strip out failed operations
-
Reduce over a history in as few passes as possible, computing auxiliary data structures
-
Partition a history into several histories, one per key, and analyze each partition separately.
-
Perform a series of reductions over a history, ideally in parallel, using data from each reduction stage to influence the second stage, and possibly a third, etc.
Overall this library aims for a small-but-useful footprint which de-duplicates analysis routines that are often split up between Jepsen itself and checkers like Knossos.
Operations in a history are always represented by defrecords, which offer improved performance and map-like semantics. For ease-of-use at the REPL and when serializing to EDN, they print like normal maps.
Histories are essentially "vectors, plus some auxiliary data structures to make querying faster". They are immutable, persistent structures. They transparently wrap an ordered collection of operations and behave like vectors of operations in terms of count, nth, seq, reduce, and so on. Like other Clojure collections, their hashes and equality semantics are based purely on their elements, not their internal structure or metadata. This makes writing tests and working at the REPL easy. Histories are drop-in replacements for existing Jepsen checkers written to work on vectors or seqs.
However, unlike vectors, you can ask a history to do things like "find the invocation corresponding to this completion", and it'll do that fast.
This library tries to be reasonably fast (at least for Clojure)--supporting parallel techniques, memoization, and compact data structures.
We've recorded and analyzed histories up to a billion operations. Compatibility with extant JVM APIs imposes a theoretical upper bound of 2^31 - 1 operations, but we could conceivably go larger.
In general, analyses of Jepsen histories is done in-memory. Recent work in Jepsen has brough us to streaming-capable on-disk storage, but some checkers require materializing large chunks of the history in RAM. I'd like to explore doing analyses meaningfully without materializing the whole history in memory. This library has been written with an eye towards writing checkers that never materialize the whole dataset in RAM, and to fuse together multiple reductions into a single IO pass over the data.
Copyright © 2022, 2023 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.