Logging
Mesos handles the logs of each Mesos component differently depending on the degree of control Mesos has over the source code of the component.
Roughly, these categories are:
- Internal - Master and Agent.
- Containers - Executors and Tasks.
- External - Components launched outside of Mesos, like Frameworks and ZooKeeper. These are expected to implement their own logging solution.
Internal
The Mesos Master and Agent use the Google's logging library. For information regarding the command-line options used to configure this library, see the configuration documentation. Google logging options that are not explicitly mentioned there can be configured via environment variables.
Both Master and Agent also expose a /logging/toggle HTTP endpoint which temporarily toggles verbose logging:
POST <ip:port>/logging/toggle?level=[1|2|3]&duration=VALUE
The effect is analogous to setting the GLOG_v
environment variable prior
to starting the Master/Agent, except the logging level will revert to the
original level after the given duration.
Containers
For background, see the containerizer documentation.
Mesos does not assume any structured logging for entities running inside containers. Instead, Mesos will store the stdout and stderr of containers into plain files ("stdout" and "stderr") located inside the sandbox.
In some cases, the default Container logger behavior of Mesos is not ideal:
- Logging may not be standardized across containers.
- Logs are not easily aggregated.
- Log file sizes are not managed. Given enough time, the "stdout" and "stderr" files can fill up the Agent's disk.
ContainerLogger
Module
The ContainerLogger
module was introduced in Mesos 0.27.0 and aims to address
the shortcomings of the default logging behavior for containers. The module
can be used to change how Mesos redirects the stdout and stderr of containers.
The interface for a ContainerLogger
can be found here.
Mesos comes with two ContainerLogger
modules:
- The
SandboxContainerLogger
implements the existing logging behavior as aContainerLogger
. This is the default behavior. - The
LogrotateContainerLogger
addresses the problem of unbounded log file sizes.
LogrotateContainerLogger
The LogrotateContainerLogger
constrains the total size of a container's
stdout and stderr files. The module does this by rotating log files based
on the parameters to the module. When a log file reaches its specified
maximum size, it is renamed by appending a .N
to the end of the filename,
where N
increments each rotation. Older log files are deleted when the
specified maximum number of files is reached.
Invoking the module
The LogrotateContainerLogger
can be loaded by specifying the library
liblogrotate_container_logger.so
in the
--modules
flag when starting the Agent and by
setting the --container_logger
Agent flag to
org_apache_mesos_LogrotateContainerLogger
.
Module parameters
Key | Explanation |
---|---|
max_stdout_size /max_stderr_size
|
Maximum size, in bytes, of a single stdout/stderr log file.
When the size is reached, the file will be rotated.
|
logrotate_stdout_options /
logrotate_stderr_options
|
Additional config options to pass into logrotate for stdout.
This string will be inserted into a logrotate configuration
file. i.e. For "stdout":
/path/to/stdout { [logrotate_stdout_options] size [max_stdout_size] }NOTE: The size option will be overridden by this module.
|
environment_variable_prefix
|
Prefix for environment variables meant to modify the behavior of
the logrotate logger for the specific container being launched.
The logger will look for four prefixed environment variables in the
container's CommandInfo 's Environment :
|
launcher_dir
|
Directory path of Mesos binaries.
The LogrotateContainerLogger will find the
mesos-logrotate-logger binary under this directory.
|
logrotate_path
|
If specified, the LogrotateContainerLogger will use the
specified logrotate instead of the system's
logrotate . If logrotate is not found, then
the module will exit with an error.
|
How it works
- Every time a container starts up, the
LogrotateContainerLogger
starts up companion subprocesses of themesos-logrotate-logger
binary. - The module instructs Mesos to redirect the container's stdout/stderr
to the
mesos-logrotate-logger
. - As the container outputs to stdout/stderr,
mesos-logrotate-logger
will pipe the output into the "stdout"/"stderr" files. As the files grow,mesos-logrotate-logger
will calllogrotate
to keep the files strictly under the configured maximum size. - When the container exits,
mesos-logrotate-logger
will finish logging before exiting as well.
The LogrotateContainerLogger
is designed to be resilient across Agent
failover. If the Agent process dies, any instances of mesos-logrotate-logger
will continue to run.
Writing a Custom ContainerLogger
For basics on module writing, see the modules documentation.
There are several caveats to consider when designing a new ContainerLogger
:
- Logging by the
ContainerLogger
should be resilient to Agent failover. If the Agent process dies (which includes theContainerLogger
module), logging should continue. This is usually achieved by using subprocesses. - When containers shut down, the
ContainerLogger
is not explicitly notified. Instead, encounteringEOF
in the container's stdout/stderr signifies that the container has exited. This provides a stronger guarantee that theContainerLogger
has seen all the logs before exiting itself. - The
ContainerLogger
should not assume that containers have been launched with any specificContainerLogger
. The Agent may be restarted with a differentContainerLogger
. - Each containerizer running on an Agent uses its own
instance of the
ContainerLogger
. This means more than oneContainerLogger
may be running in a single Agent. However, each Agent will only run a single type ofContainerLogger
.