• Argument list parsing regression fix: Custom splitter is used to allow JCommander to return an empty list when an argument that takes a list of elements is given an empty string. This matches the behavior of the previous arg parsing library.



  • Added the ability to configure the executor's stop timeout, which is the maximum amount of time the executor will wait during a graceful shutdown sequence before continuing the ‘Forceful Termination’ process (see here for details).
  • Added the ability to configure the wait period after calling the graceful shutdown endpoint and the shutdown endpoint using the graceful_shutdown_wait_secs and shutdown_wait_secs fields in HttpLifecycleConfig respectively. Previously, the executor would only wait 5 seconds between steps (adding up to a total of 10 seconds as there are 2 steps). The overall waiting period is bounded by the executor‘s stop timeout, which can be configured using the executor’s stop_timeout_in_secs flag.
  • Added the thrift_method_interceptor_modules scheduler flag that lets cluster operators inject custom Thrift method interceptors.
  • Increase default ZooKeeper session timeout from 4 to 15 seconds.
  • Added option -zk_connection_timeout to control the connection timeout of ZooKeeper connections.
  • Added scheduler command line argument -hold_offers_forever, suitable for use in clusters where Aurora is the only framework. This setting disables other options such as -min_offer_hold_time, and allows the scheduler to more efficiently cache scheduling attempts.
  • The scheduler no longer uses an internal H2 database for storage.
  • There is a new Scheduler UI which, in addition to the facelift, provides the ability to inject your own custom UI components.

Deprecations and removals:

  • Removed the deprecated command line argument -zk_use_curator, removing the choice to use the legacy ZooKeeper client.
  • Removed the rewriteConfigs thrift API call in the scheduler. This was a last-ditch mechanism to modify scheduler state on the fly. It was considered extremely risky to use since its inception, and is safer to abandon due to its lack of use and likelihood for code rot.
  • Removed the Job environment validation from the command line client. Validation was moved to the the scheduler side through the allowed_job_environments option. By default allowing any of devel, test, production, and any value matching the regular expression staging[0-9]*.
  • Removed scheduler command line arguments related to the internal H2 database, which is no longer used:
    • -use_beta_db_task_store
    • -enable_db_metrics
    • -slow_query_log_threshold
    • -db_row_gc_interval
    • -db_lock_timeout
    • -db_max_active_connection_count
    • -db_max_idle_connection_count
    • -snapshot_hydrate_stores
    • -enable_h2_console



  • Update to Shiro 1.2.5



  • Update to Mesos 1.2.0. Please upgrade Aurora to 0.18 before upgrading Mesos to 1.2.0 if you rely on Mesos filesystem images.
  • Add message parameter to killTasks RPC.
  • Add prune_tasks endpoint to aurora_admin. See aurora_admin prune_tasks -h for usage information.
  • Add support for per-task volume mounts for Mesos containers to the Aurora config DSL.
  • Added the -mesos_driver flag to the scheduler with three possible options: SCHEDULER_DRIVER, V0_MESOS, V1_MESOS. The first uses the original driver and the latter two use two new drivers from libmesos. V0_MESOS uses the SCHEDULER_DRIVER under the hood and V1_MESOS uses a new HTTP API aware driver. Users that want to use the HTTP API should use V1_MESOS. Performance sensitive users should stick with the SCHEDULER_DRIVER or V0_MESOS drivers.
  • Add observer command line options to control the resource collection interval for observed tasks. See here for details.
  • Added support for reserving agents during job updates, which can substantially reduce update times in clusters with high contention for resources. Disabled by default, but can be enabled with enable_update_affinity option, and the reservation timeout can be controlled via update_affinity_reservation_hold_time.
  • Add task scp command to the CLI client for easy transferring of files to/from/between task instances. See here for details. Currently only fully supported for Mesos containers (you can copy files from the Docker container sandbox but you cannot send files to it).
  • Added ability to inject your own scheduling logic, via a lazy Guice module binding. This is an alpha-level feature and not subject to backwards compatibility considerations. You can specify your custom modules using the task_assigner_modules and preemption_slot_finder_modules options.
  • Added support for resource bin-packing via the ‘-offer_order’ option. You can choose from CPU, MEMORY, DISK, RANDOM or REVOCABLE_CPU. You can also compose secondary sorts by combining orders together: e.g. to bin-pack by CPU and MEMORY you could supply ‘CPU,MEMORY’. The current default is RANDOM, which has the strong advantage that users can (usually) relocate their tasks due to noisy neighbors or machine issues with a task restart. When you have deterministic bin-packing, they may always end up on the same agent. So be careful enabling this without proper monitoring and remediation of host failures.
  • Modified job update behavior to create new instances, then update existing instances, and then kill unwanted instances. Previously, a job update would modify each instance in the order of their instance ID.
  • Added ability to whitelist TaskStateChanges in the webhook configuration file. You can specify a list of desired TaskStateChanges(represented by their task statuses) to be sent to a configured endpoint.



  • Upgraded Mesos to 1.1.0.
  • Added a new flag --snapshot_hydrate_stores that controls which H2-backed stores to write fully hydrated into the Scheduler snapshot. Can lead to significantly lower snapshot times for large clusters if you set this flag to an empty list. Old behavior is preserved by default, but see org.apache.aurora.scheduler.storage.log.SnapshotStoreImpl for which stores we currently have duplicate writes for.
  • A task's tier is now mapped to a label on the Mesos TaskInfo proto.
  • The Aurora client is now using the Thrift binary protocol to communicate with the scheduler.
  • Introduce a new --ip option to bind the Thermos observer to a specific rather than all interfaces.
  • Fix error that prevents the scheduler from being launched with -enable_revocable_ram.
  • The Aurora Scheduler API supports volume mounts per task for the Mesos Containerizer if the scheduler is running with the -allow_container_volumes flag.
  • The executor will send SIGTERM to processes that self daemonize via double forking.
  • Resolve docker tags to concrete identifiers for DockerContainerizer, so that job configuration is immutable across restarts. The feature introduces a new {{docker.image[name][tag]}} binder that can be used in the Aurora job configuration to resolve a docker image specified by its name:tag to a concrete identifier specified by its registry/name@digest. It requires version 2 of the Docker Registry.
  • Use RUNNING state to indicate that the task is healthy and behaving as expected. Job updates can now rely purely on health checks rather than watch_secs timeout when deciding an individial instance update state, by setting watch_secs to 0. A service will remain in STARTING state util min_consecutive_successes consecutive health checks have passed.
  • The default logging output has been changed to remove line numbers and inner class information in exchange for faster logging.
  • Support the deployment of the Aurora scheduler behind HTTPS-enabled reverse proxies: By launching scheduler via -serverset_endpoint_name=https you can ensure the Aurora client will correctly discover HTTPS support via the ZooKeeper-based discovery mechanism.
  • Scheduling performance has been improved by scheduling multiple tasks per scheduling round.
  • Preemption slot search logic is modified to improve its performance.
    • Multiple reservations are made per task group per round.
    • Multiple reservations are evaluated per round.
  • New scheduler metrics are added to facilitate monitoring and performance studies (AURORA-1832).



  • Upgraded Mesos to 1.0.0. Note: as part of this upgrade we have switched from depending on the mesos.native egg for Thermos in favor of the stripped down mesos.executor egg. This means users launching Docker tasks with the Mesos DockerContainerizer are no longer required to use images that include all of Mesos's dependencies.
  • Scheduler command line behavior has been modified to warn users of the deprecation of production attribute in Job thrift struct. The scheduler is queried for tier configurations and the user's choice of tier and production attributes is revised, if necessary. If tier is already set, the production attribute might be adjusted to match the tier selection. Otherwise, tier is selected based on the value of production attribute. If a matching tier is not found, the default tier from tier configuration file (tiers.json) is used.
  • The /offers endpoint has been modified to display attributes of resource offers as received from Mesos. This has affected rendering of some of the existing attributes. Furthermore, it now dumps additional offer attributes including reservations and persistent volumes.
  • The scheduler API now accepts both thrift JSON and binary thrift. If a request is sent without a Content-Type header, or a Content-Type header of application/x-thrift or application/json or application/vnd.apache.thrift.json the request is treated as thrift JSON. If a request is sent with a Content-Type header of application/vnd.apache.thrift.binary the request is treated as binary thrift. If the Accept header of the request is application/vnd.apache.thrift.binary then the response will be binary thrift. Any other value for Accept will result in thrift JSON.
  • Scheduler is now able to launch jobs using more than one executor at a time. To use this feature the -custom_executor_config flag must point to a JSON file which contains at least one valid executor configuration as detailed in the configuration documentation.
  • Add rollback API to the scheduler and new client command to support rolling back active update jobs to their initial state.
  • The scheduler flag -zk_use_curator now defaults to true and care should be taken when upgrading from a configuration that does not pass the flag. The scheduler upgrade should be performed by bringing all schedulers down, and then bringing upgraded schedulers up. A rolling upgrade would result in no leading scheduler for the duration of the roll which could be confusing to monitor and debug.
  • A new command aurora_admin reconcile_tasks is now available on the Aurora admin client that can trigger implicit and explicit task reconciliations.
  • Add a new MTTS (Median Time To Starting) metric in addition to MTTA and MTTR.
  • In addition to CPU resources, RAM resources can now be treated as revocable via the scheduler commandline flag -enable_revocable_ram.
  • Introduce UpdateMetadata fields in JobUpdateRequest to allow clients to store metadata on update.
  • Changed cronSchedule field inside of JobConfiguration schema to be optional for compatibility with Go.
  • Update default value of command line option -framework_name to ‘Aurora’.
  • Tasks launched with filesystem images and the Mesos unified containerizer are now fully isolated from the host‘s filesystem. As such they are no longer required to include any of the executor’s dependencies (e.g. Python 2.7) within the task's filesystem.

Deprecations and removals:

  • The job configuration flag production is now deprecated. To achieve the same scheduling behavior that production=true used to provide, users should elect a tier for the job with attributes preemptible=false and revocable=false. For example, the preferred tier in the default tier configuration file (tiers.json) matches the above criteria.
  • The ExecutorInfo.source field is deprecated and has been replaced with a label named source. It will be removed from Mesos in a future release.
  • The scheduler flag -zk_use_curator has been deprecated. If you have never set the flag and are upgrading you should take care as described in the note above.
  • The key argument of getJobUpdateDetails has been deprecated. Use the query argument instead.
  • The --release-threshold option on aurora job restart has been removed.



  • New scheduler commandline argument -enable_mesos_fetcher to allow job submissions to contain URIs which will be passed to the Mesos Fetcher and subsequently downloaded into the sandbox. Please note that enabling job submissions to download resources from arbitrary URIs may have security implications.
  • Upgraded Mesos to 0.28.2.

Deprecations and removals:



  • Upgraded Mesos to 0.27.2

  • Added a new optional Apache Curator backend for performing scheduler leader election. You can enable this with the new -zk_use_curator scheduler argument.

  • Adding --nosetuid-health-checks flag to control whether the executor runs health checks as the job‘s role’s user.

  • New scheduler command line argument -offer_filter_duration to control the time after which we expect Mesos to re-offer unused resources. A short duration improves scheduling performance in smaller clusters, but might lead to resource starvation for other frameworks if you run multiple ones in your cluster. Uses the Mesos default of 5s.

  • New scheduler command line option -framework_name to change the name used for registering the Aurora framework with Mesos. The current default value is ‘TwitterScheduler’.

  • Added experimental support for launching tasks using filesystem images and the Mesos unified containerizer. See that linked documentation for details on configuring Mesos to use the unified containerizer. Note that earlier versions of Mesos do not fully support the unified containerizer. Mesos 0.28.x or later is recommended for anyone adopting task images via the Mesos containerizer.

  • Upgraded to pystachio 0.8.1 to pick up support for the new Choice type.

  • The container property of a Job is now a Choice of either a Container holder, or a direct reference to either a Docker or Mesos container.

  • New scheduler command line argument -ip to control what ip address to bind the schedulers http server to.

  • Added experimental support for Mesos GPU resource. This feature will be available in Mesos 1.0 and is disabled by default. Use -allow_gpu_resource flag to enable it.

    IMPORTANT: once this feature is enabled, creating jobs with GPU resource will make scheduler snapshot backwards incompatible. Scheduler will be unable to read snapshot if rolled back to previous version. If rollback is absolutely necessary, perform the following steps:

    1. Set -allow_gpu_resource to false
    2. Delete all jobs with GPU resource (including cron job schedules if applicable)
    3. Wait until GPU task history is pruned. You may speed it up by changing the history retention flags, e.g.: -history_prune_threshold=1mins and -history_max_per_job_threshold=0
    4. In case there were GPU job updates created, prune job update history for affected jobs from /h2console endpoint or reduce job update pruning thresholds, e.g.: -job_update_history_pruning_threshold=1mins and -job_update_history_per_job_threshold=0
    5. Ensure a new snapshot is created by running aurora_admin scheduler_snapshot <cluster>
    6. Rollback to previous version
  • Experimental support for a webhook feature which POSTs all task state changes to a user defined endpoint.

  • Added support for specifying the default tier name in tier configuration file (tiers.json). The default property is required and is initialized with the preemptible tier (preemptible tier tasks can be preempted but their resources cannot be revoked).

Deprecations and removals:

  • Deprecated --restart-threshold option in the aurora job restart command to match the job updater behavior. This option has no effect now and will be removed in the future release.
  • Deprecated -framework_name default argument ‘TwitterScheduler’. In a future release this will change to ‘aurora’. Please be aware that depending on your usage of Mesos, this will be a backward incompatible change. For details, see MESOS-703.
  • The -thermos_observer_root command line arg has been removed from the scheduler. This was a relic from the time when executor checkpoints were written globally, rather than into a task's sandbox.
  • Setting the container property of a Job to a Container holder is deprecated in favor of setting it directly to the appropriate (i.e. Docker or Mesos) container type.
  • Deprecated numCpus, ramMb and diskMb fields in TaskConfig and ResourceAggregate thrift structs. Use set<Resource> resources to specify task resources or quota values.
  • The endpoint /slaves is deprecated. Please use /agents instead.
  • Deprecated production field in TaskConfig thrift struct. Use tier field to specify task scheduling and resource handling behavior.
  • The scheduler resources_*_ram_gb and resources_*_disk_gb metrics have been renamed to resources_*_ram_mb and resources_*_disk_mb respectively. Note the unit change: GB -> MB.



  • Upgraded Mesos to 0.26.0
  • Added a new health endpoint (/leaderhealth) which can be used for load balancer health checks to always forward requests to the leading scheduler.
  • Added a new aurora job add client command to scale out an existing job.
  • Upgraded the scheduler ZooKeeper client from 3.4.6 to 3.4.8.
  • Added support for dedicated constraints not exclusive to a particular role. See here for more details.
  • Added a new argument --announcer-hostname to thermos executor to override hostname in service registry endpoint. See here for details.
  • Descheduling a cron job that was not actually scheduled will no longer return an error.
  • Added a new argument -thermos_home_in_sandbox to the scheduler for optionally changing HOME to the sandbox during thermos executor/runner execution. This is useful in cases where the root filesystem inside of the container is read-only, as it moves PEX extraction into the sandbox. See here for more detail.
  • Support for ZooKeeper authentication in the executor announcer. See here for details.
  • Scheduler H2 in-memory database is now using MVStore In addition, scheduler thrift snapshots are now supporting full DB dumps for faster restarts.
  • Added scheduler argument -require_docker_use_executor that indicates whether the scheduler should accept tasks that use the Docker containerizer without an executor (experimental).
  • Jobs referencing invalid tier name will be rejected by the scheduler.
  • Added a new scheduler argument --populate_discovery_info. If set to true, Aurora will start to populate DiscoveryInfo field on TaskInfo of Mesos. This could be used for alternative service discovery solution like Mesos-DNS.
  • Added support for automatic schema upgrades and downgrades when restoring a snapshot that contains a DB dump.

Deprecations and removals:

  • Removed deprecated (now redundant) fields:
    • Identity.role
    • TaskConfig.environment
    • TaskConfig.jobName
    • TaskQuery.owner
  • Removed deprecated AddInstancesConfig parameter to addInstances RPC.
  • Removed deprecated executor argument -announcer-enable, which was a no-op in 0.12.0.
  • Removed deprecated API constructs related to Locks:
    • removed RPCs that managed locks
      • acquireLock
      • releaseLock
      • getLocks
    • removed Lock parameters to RPCs
      • createJob
      • scheduleCronJob
      • descheduleCronJob
      • restartShards
      • killTasks
      • addInstances
      • replaceCronTemplate
  • Task ID strings are no longer prefixed by a timestamp.
  • Changes to the way the scheduler reads command line arguments
    • Removed support for reading command line argument values from files.
    • Removed support for specifying command line argument names with fully-qualified class names.



  • Upgraded Mesos to 0.25.0.
  • Upgraded the scheduler ZooKeeper client from 3.3.4 to 3.4.6.
  • Added support for configuring Mesos role by passing -mesos_role to Aurora scheduler at start time. This enables resource reservation for Aurora when running in a shared Mesos cluster.
  • Aurora task metadata is now mapped to Mesos task labels. Labels are prefixed with org.apache.aurora.metadata. to prevent clashes with other, external label sources.
  • Added new scheduler flag -default_docker_parameters to allow a cluster operator to specify a universal set of parameters that should be used for every container that does not have parameters explicitly configured at the job level.
  • Added support for jobs to specify arbitrary ZooKeeper paths for service registration. See here for details.
  • Log destination is configurable for the thermos runner. See the configuration reference for details on how to configure destination per-process. Command line options may also be passed through the scheduler in order to configure the global default behavior.
  • Env variables can be passed through to task processes by passing --preserve_env to thermos.
  • Changed scheduler logging to use logback. Operators wishing to customize logging may do so with standard logback configuration
  • When using --read-json, aurora can now load multiple jobs from one json file, similar to the usual pystachio structure: {"jobs": [job1, job2, ...]}. The older single-job json format is also still supported.
  • aurora config list command now supports --read-json
  • Added scheduler command line argument -shiro_after_auth_filter. Optionally specify a class implementing javax.servlet.Filter that will be included in the Filter chain following the Shiro auth filters.
  • The addInstances thrift RPC does now increase job instance count (scale out) based on the task template pointed by instance key.

Deprecations and removals:

  • Deprecated AddInstancesConfig argument in addInstances thrift RPC.
  • Deprecated TaskQuery argument in killTasks thrift RPC to disallow killing tasks across multiple roles. The new safer approach is using JobKey with instances instead.
  • Removed the deprecated field ‘ConfigGroup.instanceIds’ from the API.
  • Removed the following deprecated HealthCheckConfig client-side configuration fields: endpoint, expected_response, expected_response_code. These are now set exclusively in like-named fields of HttpHealthChecker.
  • Removed the deprecated ‘JobUpdateSettings.maxWaitToInstanceRunningMs’ thrift api field ( UpdateConfig.restart_threshold in client-side configuration). This aspect of job restarts is now controlled exclusively via the client with aurora job restart --restart-threshold=[seconds].
  • Deprecated executor flag --announcer-enable. Enabling the announcer previously required both flags --announcer-enable and --announcer-ensemble, but now only --announcer-ensemble must be set. --announcer-enable is a no-op flag now and will be removed in future version.
  • Removed scheduler command line arguments:
    • -enable_cors_support. Enabling CORS is now implicit by setting the argument -enable_cors_for.
    • -deduplicate_snapshots and -deflate_snapshots. These features are good to always enable.
    • -enable_job_updates and -enable_job_creation
    • -extra_modules
    • -logtostderr, -alsologtostderr, -vlog, -vmodule, and use_glog_formatter. Removed in favor of the new logback configuration.



  • Upgraded Mesos to 0.24.1.
  • Added a new scheduler flag ‘framework_announce_principal’ to support use of authorization and rate limiting in Mesos.
  • Added support for shell-based health checkers in addition to HTTP health checkers. In concert with this change the HealthCheckConfig schema has been restructured to more cleanly allow configuring varied health checkers.
  • Added support for taking in an executor configuration in JSON via a command line argument --custom_executor_config which will override all other the command line arguments and default values pertaining to the executor.
  • Log rotation has been added to the thermos runner. See the configuration reference for details on how configure rotation per-process. Command line options may also be passed through the scheduler in order to configure the global default behavior.

Deprecations and removals:

  • The client-side updater has been removed, along with the CLI commands that used it: ‘aurora job update’ and ‘aurora job cancel-update’. Users are encouraged to take advantage of scheduler-driven updates (see ‘aurora update -h’ for usage), which has been a stable feature for several releases.
  • The following fields from HealthCheckConfig are now deprecated: endpoint, expected_response, expected_response_code in favor of setting them as part of an HttpHealthChecker.
  • The field ‘JobUpdateSettings.maxWaitToInstanceRunningMs’ (UpdateConfig.restart_threshold in client-side configuration) is now deprecated. This setting was brittle in practice, and is ignored by the 0.11.0 scheduler.



  • Upgraded Mesos to 0.23.0. NOTE: Aurora executor now requires openssl runtime dependencies that were not previously enforced. You will need libcurl available on every Mesos slave (or Docker container) to successfully launch Aurora executor. See here for more details on Mesos runtime dependencies.
  • Resource quota is no longer consumed by production jobs with a dedicated constraint (AURORA-1457).
  • The Python build layout has changed:
    • The apache.thermos package has been removed.
    • The apache.gen.aurora package has been renamed to apache.aurora.thrift.
    • The apache.gen.thermos package has been renamed to apache.thermos.thrift.
    • A new apache.thermos.runner package has been introduced, providing the thermos_runner binary.
    • A new apache.aurora.kerberos package has been introduced, containing the Kerberos-supporting versions of aurora and aurora_admin (kaurora and kaurora_admin).
    • Most BUILD targets under src/main have been removed, see here for details.

Deprecations and removals:

  • Removed the --root option from the observer.
  • Thrift ConfigGroup.instanceIds field has been deprecated. Use ConfigGroup.instances instead.
  • Deprecated SessionValidator and CapabilityValidator interfaces have been removed. All SessionKey-typed arguments are now nullable and ignored by the scheduler Thrift API.


  • Now requires JRE 8 or greater.
  • GC executor is fully replaced by the task state reconciliation (AURORA-1047).
  • The scheduler command line argument -enable_legacy_constraints has been removed, and the scheduler no longer automatically injects host and rack constraints for production services. (AURORA-1074)
  • SLA metrics for non-production jobs have been disabled by default. They can be enabled via the scheduler command line. Metric names have changed from ...nonprod_ms to ...ms_nonprod (AURORA-1350).


  • A new command line argument was added to the observer: --mesos-root This must point to the same path as --work_dir on the mesos slave.
  • Build targets for thermos and observer have changed, they are now:
    • src/main/python/apache/aurora/tools:thermos
    • src/main/python/apache/aurora/tools:thermos_observer