On Wed, Nov 6, 2019 at 11:48 AM Kyle Weaver <[email protected]> wrote:
> The way the Python SDK currently does this is to use the version as the > default tag, eg 2.16.0. While master uses 2.16.0.dev. This means there > should never be any conflicts between a release and developer image, unless > the user deliberately changes the image tags. > > > if a users' pipeline is relies on a container image released by Beam ( > or maybe a third party), external updates to such container image may not > propagate to the pipeline workflow without an explicit pull > > There should only be one released container per release. Upgrades to a > container image should not happen independently of the release process. > Fair point, although we have not yet encountered issues requiring an update of a previously released Docker, so I would not rule out considerations requiring us to re-release the image under the same tag. A scenario that is possible today is multiple pushes of container image to docker repo before the Beam release is finalized, so early adopters may be affected by stale images without a pull. > Note that so far I've just been discussing defaults. It's always possible > to use a custom container using environment_config, as mentioned earlier. > My understanding is that to pull or not to pull decision equally applies to custom image provided by environment config. > The goal is to make that unnecessary for most everyday use cases and > development. Using different container images for different transforms is a > more specialized use case worth a separate discussion. > > On Wed, Nov 6, 2019 at 11:33 AM Valentyn Tymofieiev <[email protected]> > wrote: > >> Anyway, I agree with Thomas that implicitly running `docker pull` is >>> confusing and requires some adjustments to work around. The user can always >>> run `docker pull` themselves if that's the intention. >> >> >> I understand that implicit pull may come across as surprising. However I >> see the required adjustments as a better practice. I would argue that >> customized containers images should not reuse the same name:tag >> combination, and it would also help the users avoid a situation where a >> runner may use a different container image in different execution >> environments. >> It may also help avoid issue where a user reports an issue with Beam, >> that others cannot reproduce only because a user was running a customized >> container on their local machine (and forgot about it). >> Also, if a users' pipeline is relies on a container image released by >> Beam ( or maybe a third party), external updates to such container image >> may not propagate to the pipeline workflow without an explicit pull >> >>> > 1. Read sdk version from gradle.properties and use this as the default >>> tag. Done with Python, need to implement it with Java and Go. >>> >>> 100% agree with this one. Using the same tag for local and release >>> images has already caused a good deal of confusion. Filed BEAM-8570 and >>> BEAM-8571 [2][3]. >>> >>> > 2. Remove pulling images before executing docker run command. This >>> should be fixed for Python, Java and Go. >>> >>> Valentyn (from [1]): >>> > I think pulling the latest image for the current tag is actually a >>> desired behavior, in case the external image was updated (due to a bug fix >>> for example). >>> >>> There's a PR for this [4]. Once we fix the default tag for Java/Go >>> containers, the dev and release containers will be distinct, which makes it >>> seldom important whether or not the image is `docker pull`ed. Anyway, I >>> agree with Thomas that implicitly running `docker pull` is confusing and >>> requires some adjustments to work around. The user can always run `docker >>> pull` themselves if that's the intention. >>> >>> [1] >>> https://lists.apache.org/thread.html/0f2ccbbe7969b91dc21ba331c1a30d730e268cc0355c1ac1ba0b7988@%3Cdev.beam.apache.org%3E >>> [2] https://issues.apache.org/jira/browse/BEAM-8570 >>> [3] https://issues.apache.org/jira/browse/BEAM-8571 >>> [4] https://github.com/apache/beam/pull/9972 >>> >>> On Wed, Oct 2, 2019 at 5:32 PM Ahmet Altay <[email protected]> wrote: >>> >>>> I do not believe this is a blocker for Beam 2.16. I agree that it would >>>> be good to fix this. >>>> >>>> On Wed, Oct 2, 2019 at 3:15 PM Hannah Jiang <[email protected]> >>>> wrote: >>>> >>>>> Hi Thomas >>>>> >>>>> Thanks for bring this up. >>>>> >>>>> Now Python uses sdk version as a default tag, while Java and Go use >>>>> latest as a default tag. I agree using latest as a tag is problematic. The >>>>> reason only Python uses sdk version as a default tag is Python has >>>>> version.py so the version is easy to read. For Java and Go, we need to >>>>> read >>>>> it from gradle.properties when creating images with the default tag and >>>>> when setting the default image. >>>>> >>>>> Here is what we need to do: >>>>> 1. Read sdk version from gradle.properties and use this as the default >>>>> tag. Done with Python, need to implement it with Java and Go. >>>>> 2. Remove pulling images before executing docker run command. This >>>>> should be fixed for Python, Java and Go. >>>>> >>>>> Is this a blocker for 2.16? If so and above are too much work for 2.16 >>>>> at the moment, we can hardcode the default tag for release branch for now. >>>>> >>>>> Using timestamp as a tag is an option as well, as long as runners know >>>>> which timestamp they should use. >>>>> >>>>> Hannah >>>>> >>>>> On Wed, Oct 2, 2019 at 10:13 AM Alan Myrvold <[email protected]> >>>>> wrote: >>>>> >>>>>> Yes, using the latest tag is problematic and can lead to unexpected >>>>>> behavior. >>>>>> Using a date/time or 2.17.0.dev-$USER tag would be better. The >>>>>> validates container shell script uses a datetime >>>>>> <https://github.com/apache/beam/blob/6551d0937ee31a8e310b63b222dbc750ec9331f8/sdks/python/container/run_validatescontainer.sh#L87> >>>>>> tag, which allows a unique name if no two tests are run in the same >>>>>> second. >>>>>> >>>>>> On Wed, Oct 2, 2019 at 10:05 AM Thomas Weise <[email protected]> wrote: >>>>>> >>>>>>> Want to bump this thread. >>>>>>> >>>>>>> If the current behavior is to replace locally built image with the >>>>>>> last published, then this is not only unexpected for developers but also >>>>>>> problematic for the CI, where tests should run against what was built >>>>>>> from >>>>>>> source. Or am I missing something? >>>>>>> >>>>>>> Thanks, >>>>>>> Thomas >>>>>>> >>>>>>> >>>>>>> On Tue, Sep 24, 2019 at 7:08 PM Thomas Weise <[email protected]> wrote: >>>>>>> >>>>>>>> Hi Hannah, >>>>>>>> >>>>>>>> I believe this is unexpected from the developer perspective. When >>>>>>>> building something locally, we do expect that to be used. We may need >>>>>>>> to >>>>>>>> change to not pull when the image is available locally, at least when >>>>>>>> it is >>>>>>>> a snapshot/master branch. Release images should be immutable anyways. >>>>>>>> >>>>>>>> Thomas >>>>>>>> >>>>>>>> >>>>>>>> On Tue, Sep 24, 2019 at 4:13 PM Hannah Jiang < >>>>>>>> [email protected]> wrote: >>>>>>>> >>>>>>>>> A minor update, with custom container, the pipeline would not >>>>>>>>> fail, it throws out warning and moves on to `docker run` command. >>>>>>>>> >>>>>>>>> On Tue, Sep 24, 2019 at 4:05 PM Hannah Jiang < >>>>>>>>> [email protected]> wrote: >>>>>>>>> >>>>>>>>>> Hi Brian >>>>>>>>>> >>>>>>>>>> If we pull docker images, it always downloads from remote >>>>>>>>>> repository, which is expected behavior. >>>>>>>>>> In case we want to run a local image and pull it only when the >>>>>>>>>> image is not available at local, we can use `docker run` command >>>>>>>>>> directly, >>>>>>>>>> without pulling it in advance. [1] >>>>>>>>>> In case we want to pull images only when they are not >>>>>>>>>> available at local, we can use `docker images -q` to check if images >>>>>>>>>> are >>>>>>>>>> existing at local before pulling it. >>>>>>>>>> Another option is re-tag your local image, pass your image to >>>>>>>>>> pipeline and overwrite default one, but the code is still trying to >>>>>>>>>> pull, >>>>>>>>>> so if your image is not pushed to the remote repository, it would >>>>>>>>>> fail. >>>>>>>>>> >>>>>>>>>> 1. https://github.com/docker/cli/pull/1498 >>>>>>>>>> >>>>>>>>>> Hannah >>>>>>>>>> >>>>>>>>>> On Tue, Sep 24, 2019 at 11:56 AM Brian Hulette < >>>>>>>>>> [email protected]> wrote: >>>>>>>>>> >>>>>>>>>>> I'm working on a demo cross-language pipeline on a local flink >>>>>>>>>>> cluster that relies on my python row coder PR [1]. The PR includes >>>>>>>>>>> some >>>>>>>>>>> changes to the Java worker code, so I need to build a Java SDK >>>>>>>>>>> container >>>>>>>>>>> locally and use that in the pipeline. >>>>>>>>>>> >>>>>>>>>>> Unfortunately, whenever I run the pipeline, >>>>>>>>>>> the apachebeam/java_sdk:latest tag is moved off of my locally built >>>>>>>>>>> image >>>>>>>>>>> to a newly downloaded image with a creation date 2 weeks ago, and >>>>>>>>>>> that >>>>>>>>>>> image is used instead. It looks like the reason is we run `docker >>>>>>>>>>> pull` >>>>>>>>>>> before running the container [2]. As the comment says this should >>>>>>>>>>> be a >>>>>>>>>>> no-op if the image already exists, but that doesn't seem to be the >>>>>>>>>>> case. If >>>>>>>>>>> I just run `docker pull apachebeam/java_sdk:latest` on my local >>>>>>>>>>> machine it >>>>>>>>>>> downloads the 2 week old image and happily informs me: >>>>>>>>>>> >>>>>>>>>>> Status: Downloaded newer image for apachebeam/java_sdk:latest >>>>>>>>>>> >>>>>>>>>>> Does anyone know how I can prevent `docker pull` from doing >>>>>>>>>>> this? I can unblock myself for now just by commenting out the >>>>>>>>>>> docker pull >>>>>>>>>>> command, but I'd like to understand what is going on here. >>>>>>>>>>> >>>>>>>>>>> Thanks, >>>>>>>>>>> Brian >>>>>>>>>>> >>>>>>>>>>> [1] https://github.com/apache/beam/pull/9188 >>>>>>>>>>> [2] >>>>>>>>>>> https://github.com/apache/beam/blob/master/runners/java-fn-execution/src/main/java/org/apache/beam/runners/fnexecution/environment/DockerCommand.java#L80 >>>>>>>>>>> >>>>>>>>>>
