The subunit test reporting and controlling protocol
Python C Shell Perl Makefile C++ M4
Latest commit 0de456d Jul 10, 2016 @jml jml committed on GitHub Merge pull request #20 from jeffrey4l/fix_typo
Fix typo in subunit-filter
Failed to load latest commit information.
c++ Move the C++ Listener from a patch against cppunit to a usable extern… Oct 3, 2009
filters Fix typo in subunit-filter Jul 6, 2016
m4 Add .travis.yml Jul 13, 2015
perl Update perl distcheck hack for new location. Nov 30, 2013
Apache-2.0 License change to BSD/Apache License 2.0. Sep 30, 2009
BSD License change to BSD/Apache License 2.0. Sep 30, 2009
INSTALL Handle very short packets Aug 4, 2015 Update python tarball rules for git. Aug 31, 2014 Handle very short packets Aug 4, 2015
NEWS Release 1.2. Oct 22, 2015
README.rst Typo fixes from Leo Arias Oct 22, 2015 BUG FIXES Jun 16, 2013 Release 1.2. Oct 22, 2015 Move the C++ Listener from a patch against cppunit to a usable extern… Oct 3, 2009 Switch to autoconf. May 8, 2009
setup.cfg Release 1.0.0. Nov 18, 2014 Actually ship subunit2disk. Nov 4, 2015


subunit: A streaming protocol for test results Copyright (C) 2005-2013 Robert Collins <>

Licensed under either the Apache License, Version 2.0 or the BSD 3-clause license at the users choice. A copy of both licenses are available in the project source as Apache-2.0 and BSD. You may not use this file except in compliance with one of these two licences.

Unless required by applicable law or agreed to in writing, software distributed under these licenses is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the license you chose for the specific language governing permissions and limitations under that license.

See the COPYING file for full details on the licensing of Subunit.

subunit reuses iso8601 by Michael Twomey, distributed under an MIT style licence - see python/iso8601/LICENSE for details.


Subunit is a streaming protocol for test results.

There are two major revisions of the protocol. Version 1 was trivially human readable but had significant defects as far as highly parallel testing was concerned - it had no room for doing discovery and execution in parallel, required substantial buffering when multiplexing and was fragile - a corrupt byte could cause an entire stream to be misparsed. Version 1.1 added encapsulation of binary streams which mitigated some of the issues but the core remained.

Version 2 shares many of the good characteristics of Version 1 - it can be embedded into a regular text stream (e.g. from a build system) and it still models xUnit style test execution. It also fixes many of the issues with Version 1 - Version 2 can be multiplexed without excessive buffering (in time or space), it has a well defined recovery mechanism for dealing with corrupted streams (e.g. where two processes write to the same stream concurrently, or where the stream generator suffers a bug).

More details on both protocol version s can be found in the 'Protocol' section of this document.

Subunit comes with command line filters to process a subunit stream and language bindings for python, C, C++ and shell. Bindings are easy to write for other languages.

A number of useful things can be done easily with subunit:
  • Test aggregation: Tests run separately can be combined and then reported/displayed together. For instance, tests from different languages can be shown as a seamless whole, and tests running on multiple machines can be aggregated into a single stream through a multiplexer.
  • Test archiving: A test run may be recorded and replayed later.
  • Test isolation: Tests that may crash or otherwise interact badly with each other can be run seperately and then aggregated, rather than interfering with each other or requiring an adhoc test->runner reporting protocol.
  • Grid testing: subunit can act as the necessary serialisation and deserialiation to get test runs on distributed machines to be reported in real time.
Subunit supplies the following filters:
  • tap2subunit - convert perl's TestAnythingProtocol to subunit.
  • subunit2csv - convert a subunit stream to csv.
  • subunit2disk - export a subunit stream to files on disk.
  • subunit2pyunit - convert a subunit stream to pyunit test results.
  • subunit2gtk - show a subunit stream in GTK.
  • subunit2junitxml - convert a subunit stream to JUnit's XML format.
  • subunit-diff - compare two subunit streams.
  • subunit-filter - filter out tests from a subunit stream.
  • subunit-ls - list info about tests present in a subunit stream.
  • subunit-stats - generate a summary of a subunit stream.
  • subunit-tags - add or remove tags from a stream.

Integration with other tools

Subunit's language bindings act as integration with various test runners like 'check', 'cppunit', Python's 'unittest'. Beyond that a small amount of glue (typically a few lines) will allow Subunit to be used in more sophisticated ways.


Subunit has excellent Python support: most of the filters and tools are written in python and there are facilities for using Subunit to increase test isolation seamlessly within a test suite.

The most common way is to run an existing python test suite and have it output subunit via the module:

$ python -m mypackage.tests.test_suite

For more information on the Python support Subunit offers , please see pydoc subunit, or the source in python/subunit/


Subunit has C bindings to emit the protocol. The 'check' C unit testing project has included subunit support in their project for some years now. See 'c/README' for more details.


The C library is includable and usable directly from C++. A TestListener for CPPUnit is included in the Subunit distribution. See 'c++/README' for details.


There are two sets of shell tools. There are filters, which accept a subunit stream on stdin and output processed data (or a transformed stream) on stdout.

Then there are unittest facilities similar to those for C : shell bindings consisting of simple functions to output protocol elements, and a patch for adding subunit output to the 'ShUnit' shell test runner. See 'shell/README' for details.

Filter recipes

To ignore some failing tests whose root cause is already known:

subunit-filter --without 'AttributeError.*flavor'

The xUnit test model

Subunit implements a slightly modified xUnit test model. The stock standard model is that there are tests, which have an id(), can be run, and when run start, emit an outcome (like success or failure) and then finish.

Subunit extends this with the idea of test enumeration (find out about tests a runner has without running them), tags (allow users to describe tests in ways the test framework doesn't apply any semantic value to), file attachments (allow arbitrary data to make analysing a failure easy) and timestamps.

The protocol

Version 2, or v2 is new and still under development, but is intended to supercede version 1 in the very near future. Subunit's bundled tools accept only version 2 and only emit version 2, but the new filters subunit-1to2 and subunit-2to1 can be used to interoperate with older third party libraries.

Version 2

Version 2 is a binary protocol consisting of independent packets that can be embedded in the output from tools like make - as long as each packet has no other bytes mixed in with it (which 'make -j N>1' has a tendency of doing). Version 2 is currently in draft form, and early adopters should be willing to either discard stored results (if protocol changes are made), or bulk convert them back to v1 and then to a newer edition of v2.

The protocol synchronises at the start of the stream, after a packet, or after any 0x0A byte. That is, a subunit v2 packet starts after a newline or directly after the end of the prior packet.

Subunit is intended to be transported over a reliable streaming protocol such as TCP. As such it does not concern itself with out of order delivery of packets. However, because of the possibility of corruption due to either bugs in the sender, or due to mixed up data from concurrent writes to the same fd when being embedded, subunit strives to recover reasonably gracefully from damaged data.

A key design goal for Subunit version 2 is to allow processing and multiplexing without forcing buffering for semantic correctness, as buffering tends to hide hung or otherwise misbehaving tests. That said, limited time based buffering for network efficiency is a good idea - this is ultimately implementator choice. Line buffering is also discouraged for subunit streams, as dropping into a debugger or other tool may require interactive traffic even if line buffering would not otherwise be a problem.

In version two there are two conceptual events - a test status event and a file attachment event. Events may have timestamps, and the path of multiplexers that an event is routed through is recorded to permit sending actions back to the source (such as new tests to run or stdin for driving debuggers and other interactive input). Test status events are used to enumerate tests, to report tests and test helpers as they run. Tests may have tags, used to allow tunnelling extra meanings through subunit without requiring parsing of arbitrary file attachments. Things that are not standalone tests get marked as such by setting the 'Runnable' flag to false. (For instance, individual assertions in TAP are not runnable tests, only the top level TAP test script is runnable).

File attachments are used to provide rich detail about the nature of a failure. File attachments can also be used to encapsulate stdout and stderr both during and outside tests.

Most numbers are stored in network byte order - Most Significant Byte first encoded using a variation of The first byte's top 2 high order bits encode the total number of octets in the number. This encoding can encode values from 0 to 2**30-1, enough to encode a nanosecond. Numbers that are not variable length encoded are still stored in MSB order.

prefix octets max max
00 1 2**6-1 63
01 2 2**14-1 16383
10 3 2**22-1 4194303
11 4 2**30-1 1073741823

All variable length elements of the packet are stored with a length prefix number allowing them to be skipped over for consumers that don't need to interpret them.

UTF-8 strings are with no terminating NUL and should not have any embedded NULs (implementations SHOULD validate any such strings that they process and take some remedial action (such as discarding the packet as corrupt).

In short the structure of a packet is:


In more detail...

Packets are identified by a single byte signature - 0xB3, which is never legal in a UTF-8 stream as the first byte of a character. 0xB3 starts with the first bit set and the second not, which is the UTF-8 signature for a continuation byte. 0xB3 was chosen as 0x73 ('s' in ASCII') with the top two bits replaced by the 1 and 0 for a continuation byte.

If subunit packets are being embedded in a non-UTF-8 text stream, where 0x73 is a legal character, consider either recoding the text to UTF-8, or using subunit's 'file' packets to embed the text stream in subunit, rather than the other way around.

Following the signature byte comes a 16-bit flags field, which includes a 4-bit version field - if the version is not 0x2 then the packet cannot be read. It is recommended to signal an error at this point (e.g. by emitting a synthetic error packet and returning to the top level loop to look for new packets, or exiting with an error). If recovery is desired, treat the packet signature as an opaque byte and scan for a new synchronisation point. NB: Subunit V1 and V2 packets may legitimately included 0xB3 internally, as they are an 8-bit safe container format, so recovery from this situation may involve an arbitrary number of false positives until an actual packet is encountered : and even then it may still be false, failing after passing the version check due to coincidence.

Flags are stored in network byte order too.

High byte Low byte
15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
VERSION feature bits

Valid version values are: 0x2 - version 2

Feature bits: Bit 11 - mask 0x0800 - Test id present. Bit 10 - mask 0x0400 - Routing code present. Bit 9 - mask 0x0200 - Timestamp present. Bit 8 - mask 0x0100 - Test is 'runnable'. Bit 7 - mask 0x0080 - Tags are present. Bit 6 - mask 0x0040 - File content is present. Bit 5 - mask 0x0020 - File MIME type is present. Bit 4 - mask 0x0010 - EOF marker. Bit 3 - mask 0x0008 - Must be zero in version 2.

Test status gets three bits: Bit 2 | Bit 1 | Bit 0 - mask 0x0007 - A test status enum lookup: 000 - undefined / no test 001 - Enumeration / existence 002 - In progress 003 - Success 004 - Unexpected Success 005 - Skipped 006 - Failed 007 - Expected failure

After the flags field is a number field giving the length in bytes for the entire packet including the signature and the checksum. This length must be less than 4MiB - 4194303 bytes. The encoding can obviously record a larger number but one of the goals is to avoid requiring large buffers, or causing large latency in the packet forward/processing pipeline. Larger file attachments can be communicated in multiple packets, and the overhead in such a 4MiB packet is approximately 0.2%.

The rest of the packet is a series of optional features as specified by the set feature bits in the flags field. When absent they are entirely absent.

Forwarding and multiplexing of packets can be done without interpreting the remainder of the packet until the routing code and checksum (which are both at the end of the packet). Additionally, routers can often avoid copying or moving the bulk of the packet, as long as the routing code size increase doesn't force the length encoding to take up a new byte (which will only happen to packets less than or equal to 16KiB in length) - large packets are very efficient to route.

Timestamp when present is a 32 bit unsigned integer for seconds, and a variable length number for nanoseconds, representing UTC time since Unix Epoch in seconds and nanoseconds.

Test id when present is a UTF-8 string. The test id should uniquely identify runnable tests such that they can be selected individually. For tests and other actions which cannot be individually run (such as test fixtures/layers/subtests) uniqueness is not required (though being human meaningful is highly recommended).

Tags when present is a length prefixed vector of UTF-8 strings, one per tag. There are no restrictions on tag content (other than the restrictions on UTF-8 strings in subunit in general). Tags have no ordering.

When a MIME type is present, it defines the MIME type for the file across all packets same file (routing code + testid + name uniquely identifies a file, reset when EOF is flagged). If a file never has a MIME type set, it should be treated as application/octet-stream.

File content when present is a UTF-8 string for the name followed by the length in bytes of the content, and then the content octets.

If present routing code is a UTF-8 string. The routing code is used to determine which test backend a test was running on when doing data analysis, and to route stdin to the test process if interaction is required.

Multiplexers SHOULD add a routing code if none is present, and prefix any existing routing code with a routing code ('/' separated) if one is already present. For example, a multiplexer might label each stream it is multiplexing with a simple ordinal ('0', '1' etc), and given an incoming packet with route code '3' from stream '0' would adjust the route code when forwarding the packet to be '0/3'.

Following the end of the packet is a CRC-32 checksum of the contents of the packet including the signature.

Example packets

Trivial test "foo" enumeration packet, with test id, runnable set, status=enumeration. Spaces below are to visually break up signature / flags / length / testid / crc32

b3 2901 0c 03666f6f 08555f1b

Version 1 (and 1.1)

Version 1 (and 1.1) are mostly human readable protocols.

Sample subunit wire contents

The following:

test: test foo works
success: test foo works
test: tar a file.
failure: tar a file. [
 ]..  space is eaten.
foo.c:34 WARNING foo is not defined.
a writeln to stdout

When run through subunit2pyunit:

a writeln to stdout

FAILURE: tar a file.
]..  space is eaten.
foo.c:34 WARNING foo is not defined.

Subunit v1 protocol description

This description is being ported to an EBNF style. Currently its only partly in that style, but should be fairly clear all the same. When in doubt, refer the source (and ideally help fix up the description!). Generally the protocol is line orientated and consists of either directives and their parameters, or when outside a DETAILS region unexpected lines which are not interpreted by the parser - they should be forwarded unaltered:

test|testing|test:|testing: test LABEL
success|success:|successful|successful: test LABEL
success|success:|successful|successful: test LABEL DETAILS
failure: test LABEL
failure: test LABEL DETAILS
error: test LABEL
error: test LABEL DETAILS
skip[:] test LABEL
skip[:] test LABEL DETAILS
xfail[:] test LABEL
xfail[:] test LABEL DETAILS
uxsuccess[:] test LABEL
uxsuccess[:] test LABEL DETAILS
progress: [+|-]X
progress: push
progress: pop
tags: [-]TAG ...

BRACKETED ::= '[' CR UTF8-lines ']' CR
MULTIPART ::= '[ multipart' CR PART* ']' CR
PART_TYPE ::= Content-Type: type/sub-type(;parameter=value,parameter=value)

unexpected output on stdout -> stdout. exit w/0 or last test completing -> error

Tags given outside a test are applied to all following tests Tags given after a test: line and before the result line for the same test apply only to that test, and inherit the current global tags. A '-' before a tag is used to remove tags - e.g. to prevent a global tag applying to a single test, or to cancel a global tag.

The progress directive is used to provide progress information about a stream so that stream consumer can provide completion estimates, progress bars and so on. Stream generators that know how many tests will be present in the stream should output "progress: COUNT". Stream filters that add tests should output "progress: +COUNT", and those that remove tests should output "progress: -COUNT". An absolute count should reset the progress indicators in use - it indicates that two separate streams from different generators have been trivially concatenated together, and there is no knowledge of how many more complete streams are incoming. Smart concatenation could scan each stream for their count and sum them, or alternatively translate absolute counts into relative counts inline. It is recommended that outputters avoid absolute counts unless necessary. The push and pop directives are used to provide local regions for progress reporting. This fits with hierarchically operating test environments - such as those that organise tests into suites - the top-most runner can report on the number of suites, and each suite surround its output with a (push, pop) pair. Interpreters should interpret a pop as also advancing the progress of the restored level by one step. Encountering progress directives between the start and end of a test pair indicates that a previous test was interrupted and did not cleanly terminate: it should be implicitly closed with an error (the same as when a stream ends with no closing test directive for the most recently started test).

The time directive acts as a clock event - it sets the time for all future events. The value should be a valid ISO8601 time.

The skip, xfail and uxsuccess outcomes are not supported by all testing environments. In Python the testttools ( library is used to translate these automatically if an older Python version that does not support them is in use. See the testtools documentation for the translation policy.

skip is used to indicate a test was discovered but not executed. xfail is used to indicate a test that errored in some expected fashion (also know as "TODO" tests in some frameworks). uxsuccess is used to indicate and unexpected success where a test though to be failing actually passes. It is complementary to xfail.

Hacking on subunit


  • Update versions in and python/subunit/
  • Update NEWS.
  • Do a make distcheck, which will update Makefile etc.
  • Do a PyPI release: PYTHONPATH=../../python python ../../ sdist bdist_wheel upload -s
  • Upload the regular one to LP.
  • Push a tagged commit.