2 subunit: A streaming protocol for test results
3 Copyright (C) 2005-2013 Robert Collins <robertc@robertcollins.net>
5 Licensed under either the Apache License, Version 2.0 or the BSD 3-clause
6 license at the users choice. A copy of both licenses are available in the
7 project source as Apache-2.0 and BSD. You may not use this file except in
8 compliance with one of these two licences.
10 Unless required by applicable law or agreed to in writing, software
11 distributed under these licenses is distributed on an "AS IS" BASIS, WITHOUT
12 WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
13 license you chose for the specific language governing permissions and
14 limitations under that license.
16 See the COPYING file for full details on the licensing of Subunit.
18 subunit reuses iso8601 by Michael Twomey, distributed under an MIT style
19 licence - see python/iso8601/LICENSE for details.
24 Subunit is a streaming protocol for test results.
26 There are two major revisions of the protocol. Version 1 was trivially human
27 readable but had significant defects as far as highly parallel testing was
28 concerned - it had no room for doing discovery and execution in parallel,
29 required substantial buffering when multiplexing and was fragile - a corrupt
30 byte could cause an entire stream to be misparsed. Version 1.1 added
31 encapsulation of binary streams which mitigated some of the issues but the
34 Version 2 shares many of the good characteristics of Version 1 - it can be
35 embedded into a regular text stream (e.g. from a build system) and it still
36 models xUnit style test execution. It also fixes many of the issues with
37 Version 1 - Version 2 can be multiplexed without excessive buffering (in
38 time or space), it has a well defined recovery mechanism for dealing with
39 corrupted streams (e.g. where two processes write to the same stream
40 concurrently, or where the stream generator suffers a bug).
42 More details on both protocol version s can be found in the 'Protocol' section
45 Subunit comes with command line filters to process a subunit stream and
46 language bindings for python, C, C++ and shell. Bindings are easy to write
49 A number of useful things can be done easily with subunit:
50 * Test aggregation: Tests run separately can be combined and then
51 reported/displayed together. For instance, tests from different languages
52 can be shown as a seamless whole, and tests running on multiple machines
53 can be aggregated into a single stream through a multiplexer.
54 * Test archiving: A test run may be recorded and replayed later.
55 * Test isolation: Tests that may crash or otherwise interact badly with each
56 other can be run seperately and then aggregated, rather than interfering
57 with each other or requiring an adhoc test->runner reporting protocol.
58 * Grid testing: subunit can act as the necessary serialisation and
59 deserialiation to get test runs on distributed machines to be reported in
62 Subunit supplies the following filters:
63 * tap2subunit - convert perl's TestAnythingProtocol to subunit.
64 * subunit2csv - convert a subunit stream to csv.
65 * subunit2pyunit - convert a subunit stream to pyunit test results.
66 * subunit2gtk - show a subunit stream in GTK.
67 * subunit2junitxml - convert a subunit stream to JUnit's XML format.
68 * subunit-diff - compare two subunit streams.
69 * subunit-filter - filter out tests from a subunit stream.
70 * subunit-ls - list info about tests present in a subunit stream.
71 * subunit-stats - generate a summary of a subunit stream.
72 * subunit-tags - add or remove tags from a stream.
74 Integration with other tools
75 ----------------------------
77 Subunit's language bindings act as integration with various test runners like
78 'check', 'cppunit', Python's 'unittest'. Beyond that a small amount of glue
79 (typically a few lines) will allow Subunit to be used in more sophisticated
85 Subunit has excellent Python support: most of the filters and tools are written
86 in python and there are facilities for using Subunit to increase test isolation
87 seamlessly within a test suite.
89 The most common way is to run an existing python test suite and have it output
90 subunit via the ``subunit.run`` module::
92 $ python -m subunit.run mypackage.tests.test_suite
94 For more information on the Python support Subunit offers , please see
95 ``pydoc subunit``, or the source in ``python/subunit/``
100 Subunit has C bindings to emit the protocol. The 'check' C unit testing project
101 has included subunit support in their project for some years now. See
102 'c/README' for more details.
107 The C library is includable and usable directly from C++. A TestListener for
108 CPPUnit is included in the Subunit distribution. See 'c++/README' for details.
113 There are two sets of shell tools. There are filters, which accept a subunit
114 stream on stdin and output processed data (or a transformed stream) on stdout.
116 Then there are unittest facilities similar to those for C : shell bindings
117 consisting of simple functions to output protocol elements, and a patch for
118 adding subunit output to the 'ShUnit' shell test runner. See 'shell/README' for
124 To ignore some failing tests whose root cause is already known::
126 subunit-filter --without 'AttributeError.*flavor'
132 Subunit implements a slightly modified xUnit test model. The stock standard
133 model is that there are tests, which have an id(), can be run, and when run
134 start, emit an outcome (like success or failure) and then finish.
136 Subunit extends this with the idea of test enumeration (find out about tests
137 a runner has without running them), tags (allow users to describe tests in
138 ways the test framework doesn't apply any semantic value to), file attachments
139 (allow arbitrary data to make analysing a failure easy) and timestamps.
144 Version 2, or v2 is new and still under development, but is intended to
145 supercede version 1 in the very near future. Subunit's bundled tools accept
146 only version 2 and only emit version 2, but the new filters subunit-1to2 and
147 subunit-2to1 can be used to interoperate with older third party libraries.
152 Version 2 is a binary protocol consisting of independent packets that can be
153 embedded in the output from tools like make - as long as each packet has no
154 other bytes mixed in with it (which 'make -j N>1' has a tendency of doing).
155 Version 2 is currently in draft form, and early adopters should be willing
156 to either discard stored results (if protocol changes are made), or bulk
157 convert them back to v1 and then to a newer edition of v2.
159 The protocol synchronises at the start of the stream, after a packet, or
160 after any 0x0A byte. That is, a subunit v2 packet starts after a newline or
161 directly after the end of the prior packet.
163 Subunit is intended to be transported over a reliable streaming protocol such
164 as TCP. As such it does not concern itself with out of order delivery of
165 packets. However, because of the possibility of corruption due to either
166 bugs in the sender, or due to mixed up data from concurrent writes to the same
167 fd when being embedded, subunit strives to recover reasonably gracefully from
170 A key design goal for Subunit version 2 is to allow processing and multiplexing
171 without forcing buffering for semantic correctness, as buffering tends to hide
172 hung or otherwise misbehaving tests. That said, limited time based buffering
173 for network efficiency is a good idea - this is ultimately implementator
174 choice. Line buffering is also discouraged for subunit streams, as dropping
175 into a debugger or other tool may require interactive traffic even if line
176 buffering would not otherwise be a problem.
178 In version two there are two conceptual events - a test status event and a file
179 attachment event. Events may have timestamps, and the path of multiplexers that
180 an event is routed through is recorded to permit sending actions back to the
181 source (such as new tests to run or stdin for driving debuggers and other
182 interactive input). Test status events are used to enumerate tests, to report
183 tests and test helpers as they run. Tests may have tags, used to allow
184 tunnelling extra meanings through subunit without requiring parsing of
185 arbitrary file attachments. Things that are not standalone tests get marked
186 as such by setting the 'Runnable' flag to false. (For instance, individual
187 assertions in TAP are not runnable tests, only the top level TAP test script
190 File attachments are used to provide rich detail about the nature of a failure.
191 File attachments can also be used to encapsulate stdout and stderr both during
194 Most numbers are stored in network byte order - Most Significant Byte first
195 encoded using a variation of http://www.dlugosz.com/ZIP2/VLI.html. The first
196 byte's top 2 high order bits encode the total number of octets in the number.
197 This encoding can encode values from 0 to 2**30-1, enough to encode a
198 nanosecond. Numbers that are not variable length encoded are still stored in
201 prefix octets max max
202 +-------+--------+---------+------------+
203 | 00 | 1 | 2**6-1 | 63 |
204 | 01 | 2 | 2**14-1 | 16383 |
205 | 10 | 3 | 2**22-1 | 4194303 |
206 | 11 | 4 | 2**30-1 | 1073741823 |
207 +-------+--------+---------+------------+
209 All variable length elements of the packet are stored with a length prefix
210 number allowing them to be skipped over for consumers that don't need to
213 UTF-8 strings are with no terminating NUL and should not have any embedded NULs
214 (implementations SHOULD validate any such strings that they process and take
215 some remedial action (such as discarding the packet as corrupt).
217 In short the structure of a packet is:
218 PACKET := SIGNATURE FLAGS PACKET_LENGTH TIMESTAMP? TESTID? TAGS? MIME?
219 FILECONTENT? ROUTING_CODE? CRC32
223 Packets are identified by a single byte signature - 0xB3, which is never legal
224 in a UTF-8 stream as the first byte of a character. 0xB3 starts with the first
225 bit set and the second not, which is the UTF-8 signature for a continuation
226 byte. 0xB3 was chosen as 0x73 ('s' in ASCII') with the top two bits replaced by
227 the 1 and 0 for a continuation byte.
229 If subunit packets are being embedded in a non-UTF-8 text stream, where 0x73 is
230 a legal character, consider either recoding the text to UTF-8, or using
231 subunit's 'file' packets to embed the text stream in subunit, rather than the
234 Following the signature byte comes a 16-bit flags field, which includes a
235 4-bit version field - if the version is not 0x2 then the packet cannot be
236 read. It is recommended to signal an error at this point (e.g. by emitting
237 a synthetic error packet and returning to the top level loop to look for
238 new packets, or exiting with an error). If recovery is desired, treat the
239 packet signature as an opaque byte and scan for a new synchronisation point.
240 NB: Subunit V1 and V2 packets may legitimately included 0xB3 internally,
241 as they are an 8-bit safe container format, so recovery from this situation
242 may involve an arbitrary number of false positives until an actual packet
243 is encountered : and even then it may still be false, failing after passing
244 the version check due to coincidence.
246 Flags are stored in network byte order too.
247 +-------------------------+------------------------+
248 | High byte | Low byte |
249 | 15 14 13 12 11 10 9 8 | 7 6 5 4 3 2 1 0 |
250 | VERSION |feature bits| |
251 +------------+------------+------------------------+
253 Valid version values are:
257 Bit 11 - mask 0x0800 - Test id present.
258 Bit 10 - mask 0x0400 - Routing code present.
259 Bit 9 - mask 0x0200 - Timestamp present.
260 Bit 8 - mask 0x0100 - Test is 'runnable'.
261 Bit 7 - mask 0x0080 - Tags are present.
262 Bit 6 - mask 0x0040 - File content is present.
263 Bit 5 - mask 0x0020 - File MIME type is present.
264 Bit 4 - mask 0x0010 - EOF marker.
265 Bit 3 - mask 0x0008 - Must be zero in version 2.
267 Test status gets three bits:
268 Bit 2 | Bit 1 | Bit 0 - mask 0x0007 - A test status enum lookup:
269 000 - undefined / no test
270 001 - Enumeration / existence
273 004 - Unexpected Success
276 007 - Expected failure
278 After the flags field is a number field giving the length in bytes for the
279 entire packet including the signature and the checksum. This length must
280 be less than 4MiB - 4194303 bytes. The encoding can obviously record a larger
281 number but one of the goals is to avoid requiring large buffers, or causing
282 large latency in the packet forward/processing pipeline. Larger file
283 attachments can be communicated in multiple packets, and the overhead in such a
284 4MiB packet is approximately 0.2%.
286 The rest of the packet is a series of optional features as specified by the set
287 feature bits in the flags field. When absent they are entirely absent.
289 Forwarding and multiplexing of packets can be done without interpreting the
290 remainder of the packet until the routing code and checksum (which are both at
291 the end of the packet). Additionally, routers can often avoid copying or moving
292 the bulk of the packet, as long as the routing code size increase doesn't force
293 the length encoding to take up a new byte (which will only happen to packets
294 less than or equal to 16KiB in length) - large packets are very efficient to
297 Timestamp when present is a 32 bit unsigned integer for secnods, and a variable
298 length number for nanoseconds, representing UTC time since Unix Epoch in
299 seconds and nanoseconds.
301 Test id when present is a UTF-8 string. The test id should uniquely identify
302 runnable tests such that they can be selected individually. For tests and other
303 actions which cannot be individually run (such as test
304 fixtures/layers/subtests) uniqueness is not required (though being human
305 meaningful is highly recommended).
307 Tags when present is a length prefixed vector of UTF-8 strings, one per tag.
308 There are no restrictions on tag content (other than the restrictions on UTF-8
309 strings in subunit in general). Tags have no ordering.
311 When a MIME type is present, it defines the MIME type for the file across all
312 packets same file (routing code + testid + name uniquely identifies a file,
313 reset when EOF is flagged). If a file never has a MIME type set, it should be
314 treated as application/octet-stream.
316 File content when present is a UTF-8 string for the name followed by the length
317 in bytes of the content, and then the content octets.
319 If present routing code is a UTF-8 string. The routing code is used to
320 determine which test backend a test was running on when doing data analysis,
321 and to route stdin to the test process if interaction is required.
323 Multiplexers SHOULD add a routing code if none is present, and prefix any
324 existing routing code with a routing code ('/' separated) if one is already
325 present. For example, a multiplexer might label each stream it is multiplexing
326 with a simple ordinal ('0', '1' etc), and given an incoming packet with route
327 code '3' from stream '0' would adjust the route code when forwarding the packet
330 Following the end of the packet is a CRC-32 checksum of the contents of the
331 packet including the signature.
336 Trivial test "foo" enumeration packet, with test id, runnable set,
337 status=enumeration. Spaces below are to visually break up signature / flags /
338 length / testid / crc32
340 b3 2901 0c 03666f6f 08555f1b
346 Version 1 (and 1.1) are mostly human readable protocols.
348 Sample subunit wire contents
349 ----------------------------
353 success: test foo works.
355 failure: tar a file. [
358 foo.c:34 WARNING foo is not defined.
362 When run through subunit2pyunit::
366 ========================
371 foo.c:34 WARNING foo is not defined.
374 Subunit protocol description
375 ============================
377 This description is being ported to an EBNF style. Currently its only partly in
378 that style, but should be fairly clear all the same. When in doubt, refer the
379 source (and ideally help fix up the description!). Generally the protocol is
380 line orientated and consists of either directives and their parameters, or
381 when outside a DETAILS region unexpected lines which are not interpreted by
382 the parser - they should be forwarded unaltered.
384 test|testing|test:|testing: test LABEL
385 success|success:|successful|successful: test LABEL
386 success|success:|successful|successful: test LABEL DETAILS
388 failure: test LABEL DETAILS
390 error: test LABEL DETAILS
392 skip[:] test LABEL DETAILS
394 xfail[:] test LABEL DETAILS
395 uxsuccess[:] test LABEL
396 uxsuccess[:] test LABEL DETAILS
401 time: YYYY-MM-DD HH:MM:SSZ
405 DETAILS ::= BRACKETED | MULTIPART
406 BRACKETED ::= '[' CR UTF8-lines ']' CR
407 MULTIPART ::= '[ multipart' CR PART* ']' CR
408 PART ::= PART_TYPE CR NAME CR PART_BYTES CR
409 PART_TYPE ::= Content-Type: type/sub-type(;parameter=value,parameter=value)
410 PART_BYTES ::= (DIGITS CR LF BYTE{DIGITS})* '0' CR LF
412 unexpected output on stdout -> stdout.
413 exit w/0 or last test completing -> error
415 Tags given outside a test are applied to all following tests
416 Tags given after a test: line and before the result line for the same test
417 apply only to that test, and inherit the current global tags.
418 A '-' before a tag is used to remove tags - e.g. to prevent a global tag
419 applying to a single test, or to cancel a global tag.
421 The progress directive is used to provide progress information about a stream
422 so that stream consumer can provide completion estimates, progress bars and so
423 on. Stream generators that know how many tests will be present in the stream
424 should output "progress: COUNT". Stream filters that add tests should output
425 "progress: +COUNT", and those that remove tests should output
426 "progress: -COUNT". An absolute count should reset the progress indicators in
427 use - it indicates that two separate streams from different generators have
428 been trivially concatenated together, and there is no knowledge of how many
429 more complete streams are incoming. Smart concatenation could scan each stream
430 for their count and sum them, or alternatively translate absolute counts into
431 relative counts inline. It is recommended that outputters avoid absolute counts
432 unless necessary. The push and pop directives are used to provide local regions
433 for progress reporting. This fits with hierarchically operating test
434 environments - such as those that organise tests into suites - the top-most
435 runner can report on the number of suites, and each suite surround its output
436 with a (push, pop) pair. Interpreters should interpret a pop as also advancing
437 the progress of the restored level by one step. Encountering progress
438 directives between the start and end of a test pair indicates that a previous
439 test was interrupted and did not cleanly terminate: it should be implicitly
440 closed with an error (the same as when a stream ends with no closing test
441 directive for the most recently started test).
443 The time directive acts as a clock event - it sets the time for all future
444 events. The value should be a valid ISO8601 time.
446 The skip, xfail and uxsuccess outcomes are not supported by all testing
447 environments. In Python the testttools (https://launchpad.net/testtools)
448 library is used to translate these automatically if an older Python version
449 that does not support them is in use. See the testtools documentation for the
452 skip is used to indicate a test was discovered but not executed. xfail is used
453 to indicate a test that errored in some expected fashion (also know as "TODO"
454 tests in some frameworks). uxsuccess is used to indicate and unexpected success
455 where a test though to be failing actually passes. It is complementary to
464 * Update versions in configure.ac and python/subunit/__init__.py.
465 * Make PyPI and regular tarball releases. Upload the regular one to LP, the
467 * Push a tagged commit.