Message ID | pull.31.v2.git.gitgitgadget@gmail.com (mailing list archive) |
---|---|
Headers | show |
Series | Offer to run CI/PR builds in Azure Pipelines | expand |
Hi Johannes, Thanks for putting this together, and offering to build Git on Azure Pipelines. I haven't followed v1 of this series very closely, so please excuse me if my comments have already been addressed, and I missed them in a skim of the last revision. On Mon, Oct 15, 2018 at 03:11:57AM -0700, Johannes Schindelin via GitGitGadget wrote: > It is also an invaluable tool for contributors who can validate their code > contributions via PRs on GitHub, e.g. to verify that their tests do actually > run on macOS (i.e. with the BSD family of Unix tools instead of the GNU > one). Agree. > The one sad part about this is the Windows support. Travis lacks it, and we > work around that by using Azure Pipelines (the CI part of Azure DevOps, > formerly known as Visual Studio Team Services) indirectly: one phase in > Travis would trigger a build, wait for its log, and then paste that log. I wonder if Travis' recent announcement [1] affects this at all. To summarize [1], Travis is offering an early version of adding Windows to their list of supported builder operations systems. This brings the list to macOS, Linux, and Windows, which I think satisfies what we would like to regularly build git.git on. Would we like to abandon Travis as our main CI service for upstream git.git, and build on Azure Pipelines only? If so, I think that this is an OK way to go, but I think that I would be opposed to having more than one build system. (FWIW, we tend to _three_ for Git LFS, and it can be a hassle at times). I see some benefit to sticking with Travis, since we already have a build configuration that works there. But, you've done the work to "port" that build configuration over to Azure, so perhaps the point is moot. > As Git's Windows builds (and tests!) take quite a bit of time, Travis often > timed out, or somehow the trigger did not work, and for security reasons > (the Windows builds are performed in a private pool of containers), the > Windows builds are completely disabled for Pull Requests on GitHub. This would be a concession of [1], in my mind: is it possible to run the tests on Windows in a time such that Travis will not time out? > As a special treat, this patch series adds the ability to present the > outcome of Git's test suite as JUnit-style .xml files. This allows the Azure > Pipelines build to present fun diagrams, trends, and makes it a lot easier > to drill down to test failures than before. See for example > https://dev.azure.com/git/git/_build/results?buildId=113&view=ms.vss-test-web.test-result-details > [https://dev.azure.com/git/git/_build/results?buildId=113&view=ms.vss-test-web.test-result-details] > (you can click on the label of the failed test, and then see the detailed > output in the right pane). That's pretty cool. Travis doesn't support this (to the best of my knowledge). Thanks, Taylor [1]: https://blog.travis-ci.com/2018-10-11-windows-early-release
Hi Taylor, On Mon, 15 Oct 2018, Taylor Blau wrote: > Thanks for putting this together, and offering to build Git on Azure > Pipelines. I haven't followed v1 of this series very closely, so please > excuse me if my comments have already been addressed, and I missed them > in a skim of the last revision. > > On Mon, Oct 15, 2018 at 03:11:57AM -0700, Johannes Schindelin via GitGitGadget wrote: > > It is also an invaluable tool for contributors who can validate their code > > contributions via PRs on GitHub, e.g. to verify that their tests do actually > > run on macOS (i.e. with the BSD family of Unix tools instead of the GNU > > one). > > Agree. > > > The one sad part about this is the Windows support. Travis lacks it, and we > > work around that by using Azure Pipelines (the CI part of Azure DevOps, > > formerly known as Visual Studio Team Services) indirectly: one phase in > > Travis would trigger a build, wait for its log, and then paste that log. > > I wonder if Travis' recent announcement [1] affects this at all. :-) It did not escape my notice that after years and years of trying to get *anybody* at Travis to listen to my offers to help get this started, the announcement of Azure Pipelines for OSS seemed to finally do the trick (they still don't bother to talk to me, of course). And to answer your question without a question mark: I do not really think that the Travis announcement affects this here patch series: I have a ton of good experience with Azure Pipelines, use it in Git for Windows for ages, and I am finally able to have it in core Git, too. So I want it, and I spent a lot of time getting there, and I think it probably won't hurt core Git at all (besides, it seems that at least some of the phases are a bit faster on Azure Pipelines than Travis). Another really good reason for me to do this is that I can prod the Azure Pipelines team directly. And I even get an answer, usually within minutes. Which is a lot faster than the Travis team answers my questions, which is... not yet? (I tried to get in contact with them in late 2015 or early 2016, and I tried again a year later, and then a couple of months later, and I have yet to hear back.) Also, I am not quite sure about the timeouts on Travis, but at least AppVeyor had rather short timeouts: the Windows build (due to our extensive use of Unix shell scripting in our test suite) takes 1h40m currently, and AppVeyor times out after 20 or 30 minutes. I could imagine that Travis times out after the same time, or maybe 60 minutes, which would still be too short. On Azure Pipelines, the maximum timeout (which can be configured via the azure-pipelines.yml file) is four hours IIRC. Plenty enough even for our test suite on Windows. > To summarize [1], Travis is offering an early version of adding Windows > to their list of supported builder operations systems. This brings the > list to macOS, Linux, and Windows, which I think satisfies what we would > like to regularly build git.git on. Honestly, I would love to have also FreeBSD and other platforms being tested. And with Azure Pipelines, I can make that happen (eventually), by adding another pool of VMs (given that I have a free $150/month Azure subscription, I'd use Azure VMs, of course). As long as a platform can run .NET Core software, it can run Azure Pipelines agents. With Travis, I don't think I can add private agent pools. > Would we like to abandon Travis as our main CI service for upstream > git.git, and build on Azure Pipelines only? If so, I think that this is > an OK way to go, but I think that I would be opposed to having more than > one build system. (FWIW, we tend to _three_ for Git LFS, and it can be a > hassle at times). This question of abandoning Travis in favor of Azure Pipelines is a bit of a hornets' nest, as I really, really only want to bring the goodness of Azure Pipelines to git.git, and I am *clearly* biased, as I work at Microsoft. Which is the reason why I did not even hint at it in the cover letter, let alone included a patch to make it so. My patch series is purely about adding support for running CI/PR builds of https://github.com/git/git via Azure Pipelines. > I see some benefit to sticking with Travis, since we already have a > build configuration that works there. But, you've done the work to > "port" that build configuration over to Azure, so perhaps the point is > moot. It is not so much a port, as an attempt to generalize our ci/* files. > > As Git's Windows builds (and tests!) take quite a bit of time, Travis often > > timed out, or somehow the trigger did not work, and for security reasons > > (the Windows builds are performed in a private pool of containers), the > > Windows builds are completely disabled for Pull Requests on GitHub. > > This would be a concession of [1], in my mind: is it possible to run the > tests on Windows in a time such that Travis will not time out? To be honest, I spent such a lot of time to get things to work on Azure Pipelines, *and* we get a nice view on the test failures there, too (which Travis will probably also offer soon, in response to what Azure Pipelines offer ;-)), I cannot really justify spending time on trying to make things work on Travis' Windows VMs, too. Especially when I have to expect to run into timeout issues anyway. > > As a special treat, this patch series adds the ability to present the > > outcome of Git's test suite as JUnit-style .xml files. This allows the Azure > > Pipelines build to present fun diagrams, trends, and makes it a lot easier > > to drill down to test failures than before. See for example > > https://dev.azure.com/git/git/_build/results?buildId=113&view=ms.vss-test-web.test-result-details > > [https://dev.azure.com/git/git/_build/results?buildId=113&view=ms.vss-test-web.test-result-details] > > (you can click on the label of the failed test, and then see the detailed > > output in the right pane). > > That's pretty cool. Travis doesn't support this (to the best of my > knowledge). Exactly. Plus, if things don't work in Azure Pipelines, I (or one of the other Microsoft employees among the core Git developers) can easily take a shortcut to the team and get things fixed. In my mind, that counts for a lot, too, especially given my own, frustrating personal experience with Travis. Ciao, Dscho > [1]: https://blog.travis-ci.com/2018-10-11-windows-early-release
On Mon, Oct 15, 2018 at 04:55:25PM +0200, Johannes Schindelin wrote: > Hi Taylor, > > On Mon, 15 Oct 2018, Taylor Blau wrote: > > > Thanks for putting this together, and offering to build Git on Azure > > Pipelines. I haven't followed v1 of this series very closely, so please > > excuse me if my comments have already been addressed, and I missed them > > in a skim of the last revision. > > > > On Mon, Oct 15, 2018 at 03:11:57AM -0700, Johannes Schindelin via GitGitGadget wrote: > > > It is also an invaluable tool for contributors who can validate their code > > > contributions via PRs on GitHub, e.g. to verify that their tests do actually > > > run on macOS (i.e. with the BSD family of Unix tools instead of the GNU > > > one). > > > > Agree. > > > > > The one sad part about this is the Windows support. Travis lacks it, and we > > > work around that by using Azure Pipelines (the CI part of Azure DevOps, > > > formerly known as Visual Studio Team Services) indirectly: one phase in > > > Travis would trigger a build, wait for its log, and then paste that log. > > > > I wonder if Travis' recent announcement [1] affects this at all. > > :-) > > It did not escape my notice that after years and years of trying to get > *anybody* at Travis to listen to my offers to help get this started, the > announcement of Azure Pipelines for OSS seemed to finally do the trick > (they still don't bother to talk to me, of course). > > And to answer your question without a question mark: I do not really think > that the Travis announcement affects this here patch series: I have a ton > of good experience with Azure Pipelines, use it in Git for Windows for > ages, and I am finally able to have it in core Git, too. So I want it, and > I spent a lot of time getting there, and I think it probably won't hurt > core Git at all (besides, it seems that at least some of the phases are a > bit faster on Azure Pipelines than Travis). I think that there are fair reasons to prefer Azure Pipelines over Travis. In particular, I am encouraged by the fact that we (1) know that we won't timeout, and (2) can have a standardized CI interface on Git and Git for Windows. Certainly, the Windows support on Azure Pipelines is more developed than that of Travis', so that's another point for Azure. > Another really good reason for me to do this is that I can prod the Azure > Pipelines team directly. And I even get an answer, usually within minutes. > Which is a lot faster than the Travis team answers my questions, which > is... not yet? (I tried to get in contact with them in late 2015 or early > 2016, and I tried again a year later, and then a couple of months later, > and I have yet to hear back.) Certainly a good reason. To be clear/fair, I've sent in a number of support tickets to Travis CI over the years, and have always been responded to in a short amount of time with helpful answers. I think that we would really be fine in either case, TBH. > Also, I am not quite sure about the timeouts on Travis, but at least > AppVeyor had rather short timeouts: the Windows build (due to our > extensive use of Unix shell scripting in our test suite) takes 1h40m > currently, and AppVeyor times out after 20 or 30 minutes. I could imagine > that Travis times out after the same time, or maybe 60 minutes, which > would still be too short. On Azure Pipelines, the maximum timeout (which > can be configured via the azure-pipelines.yml file) is four hours IIRC. > Plenty enough even for our test suite on Windows. > > > To summarize [1], Travis is offering an early version of adding Windows > > to their list of supported builder operations systems. This brings the > > list to macOS, Linux, and Windows, which I think satisfies what we would > > like to regularly build git.git on. > > Honestly, I would love to have also FreeBSD and other platforms being > tested. And with Azure Pipelines, I can make that happen (eventually), by > adding another pool of VMs (given that I have a free $150/month Azure > subscription, I'd use Azure VMs, of course). As long as a platform can run > .NET Core software, it can run Azure Pipelines agents. > > With Travis, I don't think I can add private agent pools. I think that's right. > > Would we like to abandon Travis as our main CI service for upstream > > git.git, and build on Azure Pipelines only? If so, I think that this is > > an OK way to go, but I think that I would be opposed to having more than > > one build system. (FWIW, we tend to _three_ for Git LFS, and it can be a > > hassle at times). > > This question of abandoning Travis in favor of Azure Pipelines is a bit of > a hornets' nest, as I really, really only want to bring the goodness of > Azure Pipelines to git.git, and I am *clearly* biased, as I work at > Microsoft. > > Which is the reason why I did not even hint at it in the cover letter, let > alone included a patch to make it so. > > My patch series is purely about adding support for running CI/PR builds of > https://github.com/git/git via Azure Pipelines. I think that adding support for one CI system does carry the weight of removing the other, for the sake of having few CI systems running at any given time. In other words, even if you don't intend to start a discussion about removing Travis, those that want run the smallest number of services (including myself) will ask you about it ;-). But I don't see a benefit to dragging this series through the mud by spending too much time talking about Travis' future. I think that your cover letter does a fine job to address the point, and we can revisit it in the future if more people than just myself are opposed to running >1 CI service. > > I see some benefit to sticking with Travis, since we already have a > > build configuration that works there. But, you've done the work to > > "port" that build configuration over to Azure, so perhaps the point is > > moot. > > It is not so much a port, as an attempt to generalize our ci/* files. > > > > As Git's Windows builds (and tests!) take quite a bit of time, Travis often > > > timed out, or somehow the trigger did not work, and for security reasons > > > (the Windows builds are performed in a private pool of containers), the > > > Windows builds are completely disabled for Pull Requests on GitHub. > > > > This would be a concession of [1], in my mind: is it possible to run the > > tests on Windows in a time such that Travis will not time out? > > To be honest, I spent such a lot of time to get things to work on Azure > Pipelines, *and* we get a nice view on the test failures there, too (which > Travis will probably also offer soon, in response to what Azure Pipelines > offer ;-)), I cannot really justify spending time on trying to make things > work on Travis' Windows VMs, too. Especially when I have to expect to run > into timeout issues anyway. Agreed. > > > As a special treat, this patch series adds the ability to present the > > > outcome of Git's test suite as JUnit-style .xml files. This allows the Azure > > > Pipelines build to present fun diagrams, trends, and makes it a lot easier > > > to drill down to test failures than before. See for example > > > https://dev.azure.com/git/git/_build/results?buildId=113&view=ms.vss-test-web.test-result-details > > > [https://dev.azure.com/git/git/_build/results?buildId=113&view=ms.vss-test-web.test-result-details] > > > (you can click on the label of the failed test, and then see the detailed > > > output in the right pane). > > > > That's pretty cool. Travis doesn't support this (to the best of my > > knowledge). > > Exactly. > > Plus, if things don't work in Azure Pipelines, I (or one of the other > Microsoft employees among the core Git developers) can easily take a > shortcut to the team and get things fixed. In my mind, that counts for a > lot, too, especially given my own, frustrating personal experience with > Travis. Agreed, and thanks for your response. > Ciao, > Dscho > > > [1]: https://blog.travis-ci.com/2018-10-11-windows-early-release Thanks, Taylor
On Mon, Oct 15, 2018 at 07:22:15AM -0700, Taylor Blau wrote: > Would we like to abandon Travis as our main CI service for upstream > git.git, and build on Azure Pipelines only? It's not only about "upstream git.git", but also about contributors, who might have enabled Travis CI integration on their forks on GitHub. Having a '.travis.yml' and associated 'ci/*' scripts in git.git makes it possible for them to easily build and test their branches on their own.
On Tue, Oct 16, 2018 at 4:55 AM Taylor Blau <me@ttaylorr.com> wrote: > > On Mon, Oct 15, 2018 at 04:55:25PM +0200, Johannes Schindelin wrote: > > Another really good reason for me to do this is that I can prod the Azure > > Pipelines team directly. And I even get an answer, usually within minutes. > > Which is a lot faster than the Travis team answers my questions, which > > is... not yet? (I tried to get in contact with them in late 2015 or early > > 2016, and I tried again a year later, and then a couple of months later, > > and I have yet to hear back.) > > Certainly a good reason. To be clear/fair, I've sent in a number of > support tickets to Travis CI over the years, and have always been > responded to in a short amount of time with helpful answers. I think > that we would really be fine in either case, TBH. Was this in the context of "I'm this random dude using Travis" or "as you guys know I work for GitHub, your biggest? customer..." ? :)
"Johannes Schindelin via GitGitGadget" <gitgitgadget@gmail.com> writes: > For a long time already, we have Git's source code continuously tested via > Travis CI, see e.g. https://travis-ci.org/git/git/builds/421738884. It has > served us well, and more and more developers actually pay attention and > benefit from the testing this gives us. What's the current status of this topic? Has the "p4 daemon gets left behind" one resolved to everybody's satisfaction? I think that one was the only large discussion on the series (aside from "do we want to keep Travis?" subthread, which does not make this series undesirable), modulo your "oy oy oy that is leftover debugging I need to remove in a reroll". The topic was marked as "On hold, monitoring discussion" and it seems that discussion has quieted down, so the next step is to see an updated series? Thanks.
Hi Junio, On Thu, 25 Oct 2018, Junio C Hamano wrote: > "Johannes Schindelin via GitGitGadget" <gitgitgadget@gmail.com> > writes: > > > For a long time already, we have Git's source code continuously tested via > > Travis CI, see e.g. https://travis-ci.org/git/git/builds/421738884. It has > > served us well, and more and more developers actually pay attention and > > benefit from the testing this gives us. > > What's the current status of this topic? Has the "p4 daemon gets > left behind" one resolved to everybody's satisfaction? No. I was kind of waiting for Luke's answer, and in the alternative I hoped to find some time to work on trying to reproduce his issues on my system (but I failed to find said time so far). > I think that one was the only large discussion on the series (aside from > "do we want to keep Travis?" subthread, which does not make this series > undesirable), modulo your "oy oy oy that is leftover debugging I need to > remove in a reroll". > > The topic was marked as "On hold, monitoring discussion" and it > seems that discussion has quieted down, so the next step is to see > an updated series? I really think that I have to figure out what causes those p4d issues before I can give you that updated. I *am* interested, to be sure, it's just that other things seem to get in my way all the time. One thing that keeps getting in my way, for example, is the performance issue identified in the chain linter. And I do think that I have to take this into consideration for another update to this here patch series, too, by adding `--no-chain-lint` to the Windows phase. There is a similar thing with `--with-dashes`, too. Will keep you updated, Dscho > > Thanks. > > >
Hi Junio, On Thu, 25 Oct 2018, Junio C Hamano wrote: > "Johannes Schindelin via GitGitGadget" <gitgitgadget@gmail.com> > writes: > > > For a long time already, we have Git's source code continuously tested via > > Travis CI, see e.g. https://travis-ci.org/git/git/builds/421738884. It has > > served us well, and more and more developers actually pay attention and > > benefit from the testing this gives us. > > What's the current status of this topic? > [...] > > The topic was marked as "On hold, monitoring discussion" and it seems > that discussion has quieted down, so the next step is to see an updated > series? See the updated series: https://public-inbox.org/git/pull.31.v3.git.gitgitgadget@gmail.com/ Thanks, Dscho
Johannes Schindelin <Johannes.Schindelin@gmx.de> writes: > ... > See the updated series: > https://public-inbox.org/git/pull.31.v3.git.gitgitgadget@gmail.com/ Thanks. I see that you are already planning for v4, but I'll find time to take a look at what is posted sometime this week anyway.
Hi Junio, On Thu, 17 Jan 2019, Junio C Hamano wrote: > Johannes Schindelin <Johannes.Schindelin@gmx.de> writes: > > > ... > > See the updated series: > > https://public-inbox.org/git/pull.31.v3.git.gitgitgadget@gmail.com/ > > Thanks. > > I see that you are already planning for v4, but I'll find time to > take a look at what is posted sometime this week anyway. Yes, v4 will ideally have only cosmetic changes apart from the part that fixes the traces of the published failed tests: -- snipsnap -- 15: f678b105f81e ! 15: 7b74987d72a6 tests: include detailed trace logs with --write-junit-xml upon failure @@ -32,6 +32,38 @@ Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de> + diff --git a/t/helper/test-path-utils.c b/t/helper/test-path-utils.c + --- a/t/helper/test-path-utils.c + +++ b/t/helper/test-path-utils.c +@@ + return !!res; + } + ++ if (argc == 4 && !strcmp(argv[1], "skip-n-bytes")) { ++ int fd = open(argv[2], O_RDONLY), offset = atoi(argv[3]); ++ char buffer[65536]; ++ ++ if (fd < 0) ++ die_errno("could not open '%s'", argv[2]); ++ if (lseek(fd, offset, SEEK_SET) < 0) ++ die_errno("could not skip %d bytes", offset); ++ for (;;) { ++ ssize_t count = read(fd, buffer, sizeof(buffer)); ++ if (count < 0) ++ die_errno("could not read '%s'", argv[2]); ++ if (!count) ++ break; ++ if (write(1, buffer, count) < 0) ++ die_errno("could not write to stdout"); ++ } ++ close(fd); ++ return 0; ++ } ++ + fprintf(stderr, "%s: unknown function name: %s\n", argv[0], + argv[1] ? argv[1] : "(there was none)"); + return 1; + diff --git a/t/test-lib.sh b/t/test-lib.sh --- a/t/test-lib.sh +++ b/t/test-lib.sh @@ -42,7 +74,8 @@ - "$(printf '%s\n' "$@" | sed 1d)")" + "$(if test -n "$GIT_TEST_TEE_OUTPUT_FILE" + then -+ cut -c "$GIT_TEST_TEE_OFFSET-" <"$GIT_TEST_TEE_OUTPUT_FILE" ++ test-tool path-utils skip-n-bytes \ ++ "$GIT_TEST_TEE_OUTPUT_FILE" $GIT_TEST_TEE_OFFSET + else + printf '%s\n' "$@" | sed 1d + fi)")" @@ -56,17 +89,17 @@ fi test_failure=$(($test_failure + 1)) @@ - write_junit_xml "$(printf '%s\n' \ - " <testcase $junit_attrs>" "$@" " </testcase>")" - junit_have_testcase=t -+ if test -n "$GIT_TEST_TEE_OUTPUT_FILE" + echo >&3 "" + maybe_teardown_valgrind + maybe_teardown_verbose ++ if test -n "$GIT_TEST_TEE_OFFSET" + then + GIT_TEST_TEE_OFFSET=$(test-tool path-utils file-size \ + "$GIT_TEST_TEE_OUTPUT_FILE") + fi } - test_done () { + test_skip () { @@ date +%Y-%m-%dT%H:%M:%S)\"" write_junit_xml --truncate "<testsuites>" " <testsuite $junit_attrs>" @@ -74,7 +107,6 @@ + if test -n "$GIT_TEST_TEE_OUTPUT_FILE" + then + GIT_TEST_TEE_OFFSET=0 -+ GIT_TEST_TEE_ERR_OFFSET=0 + fi fi