Skip to content

Latest commit

 

History

History
289 lines (227 loc) · 12.9 KB

README.md

File metadata and controls

289 lines (227 loc) · 12.9 KB
Raw

Sidecar Executor

Run Docker containers on Mesos with Sidecar service discovery! We're running it with HubSpot's Singularity scheduler.

This is a Mesos executor that integrates with the service discovery platform Sidecar to more tightly tie Sidecar into the Mesos ecosystem. The main advantage is that the executor leverages the service health checking that Sidecar already provides in order to fail Mesos tasks quickly when they have gone off the rails.

With Sidecar and Sidecar Executor you get service health checking unified with service Discovery, regardless of which Mesos scheduler you are running. The system is completely scheduler agnostic to the extent possible.

Note that unlike Sidecar, Sidecar Executor assumes that tasks are to be run as Docker containers. You may of course still integrate non-Dockerized services with Sidecar as normal.

Subset of Features

This executor does not attempt to support every single option that Docker can support for running containers. It supports the core feature set that most people actually use. If you are looking for something that it doesn't currently provide, pull requests or feature requests are welcome.

Currently supported:

  • Environment variables
  • Docker labels
  • Exposed port and port mappings
  • Volume binds from the host
  • Network mode setting
  • Capability Add
  • Capability Drop
  • Resolve environment variables stored in Vault
  • Enforce CPU and Memory limits via Docker cgroups

This set of features probably supports most of the production containers out there.

Debugability

This executor tries to expose as much information as possible to help with the painful task of debugging a Mesos task failure. It logs each of the actions that it takes and why. And, on startup it logs both the current environment variables which the executor will run with, the settings for the executor, and the Docker environment that will be supplied to the container. This can be critical in understanding how a task failed.

When a task fails or is killed, the executor will fetch the last logs that were sent to the container, and copy them to its own stdout and stderr. This means that in most cases the Mesos logs will now contain the failure messages and there is usually no need to dig further into logging frameworks to find out what happened.

Additionally since each instance of the executor manages a single container, the process name of the executor that shows up in ps output contains both the ID of the Docker container and the Docker image name that was used to start it.

Running the Executor

A separate copy of the executor is started for each task. The binary is small and only uses a few MB of memory so this is fairly cheap.

But, since the executor will always be running as long as your task, it will have the file open and you won't be ablet to replace it (e.g. to upgrade) while the task is running. It's thus recommended that you run the executor.sh script as the actual executor defined in each of your tasks. That will in turn copy the binary from a location of your choosing to the task's sandbox and then execute it from there. This means that the copy of the executor used to start the task will remain in the sandbox directory for the life of the task and solves the problem nicely.

To upgrade the executor you can then simply replace the binary in the defined location on your systems and any new tasks will start under the new copy while older tasks remain running under their respective version. By default the shell script assumes the path will be /opt/mesos/sidecar-executor.

Configuration

Each task can be run with its own executor configuration. This is helpful when in situations such as needing to delay health checking for apps that are slow to start, allowing for longer grace periods or a larger failure count before shooting a container. You may configure the executor via environment variables in the Mesos task. These will then be present for the executor at the time that it runs. Note that these are separate from the environment variables used in the Docker container. Currently the settings available are:

Setting Default
KillTaskTimeout 5 (seconds)
HttpTimeout 2s
SidecarRetryCount 5
SidecarRetryDelay 3s
SidecarUrl http://localhost:7777/state.json
SidecarBackoff 1m
SidecarPollInterval 30s
SidecarMaxFails 3
SidecarDrainingDuration 10s
SeedSidecar false
DockerRepository https://index.docker.io/v1/
LogsSince 3m
ForceCpuLimit false
ForceMemoryLimit false
UseCpuShares false
Debug false
MesosMasterPort 5050
RelaySyslog false
RelaySyslogStartupOnly false
RelaySyslogStartupTime 1m
SyslogAddr 127.0.0.1:514
ContainerLogsStdout false
SendDockerLabels []
LogHostname System Hostname

All of the environment variables are of the form EXECUTOR_SIDECAR_RETRY_DELAY where all of the CamelCased words are split apart, and each setting is prefixed with EXECUTOR_.

  • KillTaskTimeout: This is the amount of time to wait before we hard kill the container. Initially we send a SIGTERM and after this timeout we follow up with a SIGKILL. If your process has a clean shutdown procedure and takes awhile, you may want to back this off to let it complete shutdown before being hard killed by the kernel. Note that unlike the other settings, this, for internal library reasons, is an integer in seconds and not a Go time spec.

  • HttpTimeout: The timeout when talking to Sidecar. The default should be far longer than needed unless you really have something wrong.

  • SidecarRetryCount: This is the number of times we'll retry calling to Sidecar when health checking.

  • SidecarRetryDelay: The amount of time to wait between retries when contacting Sidecar.

  • SidecarUrl: The URL to use to contact Sidecar. The default will usually be the right setting.

  • SidecarBackoff: How long to wait before we start health checking to Sidecar. You want this value to be longer than the time it takes your process to start up and start responding as healthy on the health check endpoint.

  • SidecarPollInterval: The interval between asking Sidecar how healthy we are.

  • SidecarMaxFails: How many failed checks to Sidecar before we shut down the container? Note that this is not just contacting Sidecar. This is how many affirmed unhealthy checks we need to receive, each spaced apart by SidecarPollInterval.

  • SidecarDrainingDuration: How much time to wait before killing the container after instructing Sidecar to set the current service's status to DRAINING. Setting this to 0 will prevent the executor from telling Sidecar to trigger the DRAINING state and it will kill the container as soon as possible.

  • SeedSidecar: Should we query the Mesos master for the list of workers and then provide those in the SIDECAR_SEEDS environment variables?

  • DockerRepository: This is used to match the credentials that we'll store from the Docker config. This will follow the same matching order as described here. The executor expects to use only one set of credentials for each job.

  • LogsSince: When the container exits or is killed, the executor will copy logs from the Docker container output to its own stdout and stderr so that they show up in the Mesos logs. LogsSince is how far back in time we reach to get these logs.

  • ForceCpuLimit: Should we enforce the CPU limits in the request using cgroups (via Docker)?

  • ForceMemoryLimit: Should we enforce the memory limits in the request using cgroups (via Docker)?

  • UseCpuShares: By default we use the Linux Completely Fair Scheduler settings to control CPU limiting. This doesn't work well for certain workloads. Should we instead use the older CPU Shares relative workload limiting mechanism? Note that you should understand the difference before turning this on.

  • Debug: Should we turn on debug logging (verbose!) for this executor?

  • MesosMasterPort: The port on which the Mesos Master node listens on.

  • RelaySyslog: Should we relay container logs to syslog? This is a bare UDP implementation suitable for loggers that don't care about syslog protocol. Logs will be sent in JSON format, using Logrus.

  • RelaySyslogStartupOnly: Should we relay container logs to syslog only for the startup duration? This is helpful for apps that do their own syslogging but are not able to log during their startup process. The length of time to relay logs is controlled by RelaySyslogStartupTime.

  • RelaySyslogStartupTime: By itself this configuration item does nothing. It controls the value for how long to log for when RelaySyslogStartupOnly is set.

  • SyslogAddr: If RelaySyslog is true, we'll use this as the remote address for syslog logging.

  • ContainerLogsStdout: Should we copy the container logs to stdout? The effect of doing this is that container logs (both stdout and stderr) will end up in the Mesos sandbox logs. Be careful here since the Mesos logs are not rotated by the Mesos worker. Requires that RelaySyslog be true. Note that if you don't want syslog but you do want this option, there is not much harm in turning on RelaySyslog since the UDP packets will just drop. Note: This only works with Docker log drivers json-file and journald because it uses the native Docker logging functionality to collect the logs.

  • SendDockerLabels: If RelaySyslog is true, should we augment JSON logs with some fields defined in Docker labels? This is a comma-separated list of labels. They will be sent with the field name being the Docker label name.

  • LogHostname: When relaying logs, we will add this as the Hostname field. Defaults to the OS hostname and can be overridden with LOG_HOSTNAME in the environment.

Special AWS Role Configuration

In addition to all the executor level settings above, you may also pass the following env vars into the executor to enable the executor to maintain AWS creds with a specific role via Vault:

  • EXECUTOR_AWS_ROLE - specifies the AWS role (pre-existing in IAM/Vault) to request. By itself this will give you the default TTL for the policy

  • EXECUTOR_AWS_ROLE_TTL - This will allow you extend the requested time, up to the max allowed by Vault for the policy. The value is a string, specified in Go Duration format. E.g. "1m40s" for 1 minute and 40 seconds.

Vault Configuration

Because the executor uses the Vault library, it lets the Vault client configure itself from its own environment settings. You can look these up in the Vault source if you'd like to see them all. The executor adds a couple of others to better control the Vault integration. You should specify at least the following:

  • VAULT_ADDR - URL of the Vault server.
  • VAULT_MAX_RETRIES - API retries before Vault fails.
  • VAULT_TOKEN - Optional if specified in a file or using userpass.
  • VAULT_TOKEN_FILE - Where to cache Vault tokens between calls to the executor on the same host.
  • VAULT_TTL - The TTL in seconds of the Vault Token we'll have issued note that the grace period is one hour so shorter than 1 hour is not possible.

WARNING If you are using Vault, you really want to have the executor cache tokens to a VAULT_TOKEN_FILE. If not you can build up quite a lot of tokens in Vault, and that doesn't work well. Especially in a fast job failure scenario where executors running on multiple machines might be generating tokens constantly.

Configuring Docker Connectivity

Sidecar Executor supports all the normal environment variables for configuring your connection to the Docker daemon. It will look for DOCKER_HOST, etc in the runtime environment and configure connectivity accordingly.

Contributing

Contributions are more than welcome. Bug reports with specific reproduction steps are great. If you have a code contribution you'd like to make, open a pull request with suggested code.

Pull requests should:

  • Clearly state their intent in the title
  • Have a description that explains the need for the changes
  • Include tests!
  • Not break the public API

Ping us to let us know you're working on something interesting by opening a GitHub Issue on the project.