Added inspect retries to the Docker executor. This patch adds retries for `inspect` command to workaround docker daemon hangs. We assume that the docker daemon can be temporarily unresponsive. If it's unresponsive, then any started docker cli command hangs. To address the issue, we retry `inspect` in the loop.
Review: https://reviews.apache.org/r/65759/ Project: http://git-wip-us.apache.org/repos/asf/mesos/repo Commit: http://git-wip-us.apache.org/repos/asf/mesos/commit/811c5aae Tree: http://git-wip-us.apache.org/repos/asf/mesos/tree/811c5aae Diff: http://git-wip-us.apache.org/repos/asf/mesos/diff/811c5aae Branch: refs/heads/1.5.x Commit: 811c5aae799bde94ea6bceeb06303bb3e0b52d67 Parents: 65b6cbb Author: Andrei Budnik <abud...@mesosphere.com> Authored: Fri Mar 2 15:39:05 2018 -0800 Committer: Gilbert Song <songzihao1...@gmail.com> Committed: Wed Mar 7 01:08:13 2018 -0800 ---------------------------------------------------------------------- src/docker/executor.cpp | 46 +++++++++++++++++++++++++++++++++++--------- 1 file changed, 37 insertions(+), 9 deletions(-) ---------------------------------------------------------------------- http://git-wip-us.apache.org/repos/asf/mesos/blob/811c5aae/src/docker/executor.cpp ---------------------------------------------------------------------- diff --git a/src/docker/executor.cpp b/src/docker/executor.cpp index 1fb5b53..de462d9 100644 --- a/src/docker/executor.cpp +++ b/src/docker/executor.cpp @@ -22,8 +22,10 @@ #include <mesos/executor.hpp> #include <mesos/mesos.hpp> +#include <process/collect.hpp> #include <process/delay.hpp> #include <process/id.hpp> +#include <process/loop.hpp> #include <process/owned.hpp> #include <process/process.hpp> #include <process/protobuf.hpp> @@ -214,13 +216,46 @@ public: run->onAny(defer(self(), &Self::reaped, lambda::_1)); + // Since the Docker daemon might hang, we have to retry the inspect command. + auto inspectLoop = loop( + self(), + [=]() { + return await( + docker->inspect(containerName, DOCKER_INSPECT_DELAY) + .after( + DOCKER_INSPECT_TIMEOUT, + [=](Future<Docker::Container> future) { + LOG(WARNING) << "Docker inspect timed out after " + << DOCKER_INSPECT_TIMEOUT + << " for container " + << "'" << containerName << "'"; + + // We need to clean up the hanging Docker CLI process. + // Discarding the inspect future triggers a callback in + // the Docker library that kills the subprocess and + // transitions the future. + future.discard(); + return future; + })); + }, + [](const Future<Docker::Container>& future) + -> Future<ControlFlow<Docker::Container>> { + if (future.isReady()) { + return Break(future.get()); + } + if (future.isFailed()) { + return Failure(future.failure()); + } + return Continue(); + }); + // Delay sending TASK_RUNNING status update until we receive // inspect output. Note that we store a future that completes // after the sending of the running update. This allows us to // ensure that the terminal update is sent after the running // update (see `reaped()`). - inspect = docker->inspect(containerName, DOCKER_INSPECT_DELAY) - .then(defer(self(), [=](const Docker::Container& container) { + inspect = + inspectLoop.then(defer(self(), [=](const Docker::Container& container) { if (!killed) { containerPid = container.pid; @@ -321,13 +356,6 @@ public: return Nothing(); })); - inspect - .after(DOCKER_INSPECT_TIMEOUT, [=](const Future<Nothing>&) { - LOG(WARNING) << "Docker inspect has not finished after " - << DOCKER_INSPECT_TIMEOUT; - return inspect; - }); - inspect.onFailed(defer(self(), [=](const string& failure) { LOG(ERROR) << "Failed to inspect container '" << containerName << "'" << ": " << failure;
