This document serves as a guide for users who wish to upgrade an existing Mesos cluster. Some versions require particular upgrade techniques when upgrading a running cluster. Some upgrades will have incompatible changes.
Mesos 0.27 introduces the concept of implicit roles. In previous releases, configuring roles required specifying a static whitelist of valid role names on master startup (via the --roles
flag). In Mesos 0.27, if --roles
is omitted, any role name can be used; controlling which principals are allowed to register as which roles should be done using ACLs. The role whitelist functionality is still supported but is deprecated.
The Allocator API has changed due to the introduction of implicit roles. Custom allocator implementations will need to be updated. See MESOS-4000 for more information.
The executorLost
callback in the Scheduler interface will now be called whenever the slave detects termination of a custom executor. This callback was never called in previous versions, so please make sure any framework schedulers can now safely handle this callback. Note that this callback may not be reliably delivered.
The isolator prepare
interface has been changed slightly. Instead of keeping adding parameters to the prepare
interface, we decide to use a protobuf (ContainerConfig
). Also, we renamed ContainerPrepareInfo
to ContainerLaunchInfo
to better capture the purpose of this struct. See MESOS-4240 and MESOS-4282 for more information. If you are an isolator module writer, you will have to adjust your isolator module according to the new interface and re-compile with 0.27.
ACLs.shutdown_frameworks has been deprecated in favor of the new ACLs.teardown_frameworks. This affects the --acls
master flag for the local authorizer.
Reserved resources are now accounted for in the DRF role sorter. Previously unaccounted reservations will influence the weighted DRF sorter. If role weights were explicitly set, they may need to be adjusted in order to account for the reserved resources in the cluster.
NOTE The names of some TaskStatus::Reason enums have been changed. But the tag numbers remain unchanged, so it is backwards compatible. Frameworks using the new version might need to do some compile time adjustments:
NOTE The Credential
protobuf has been changed. Credential
field secret
is now a string, it used to be bytes. This will affect framework developers and language bindings ought to update their generated protobuf with the new version. This fixes JSON based credentials file support.
NOTE The /state
endpoints on master and slave will no longer include data
fields as part of the JSON models for ExecutorInfo
and TaskInfo
out of consideration for memory scalability (see MESOS-3794 and this email thread). On master, the affected data
field was originally found via frameworks[*].executors[*].data
. On slaves, the affected data
field was originally found via executors[*].tasks[*].data
.
NOTE The NetworkInfo
protobuf has been changed. The fields protocol
and ip_address
are now deprecated. The new field ip_addresses
subsumes the information provided by them.
NOTE The following endpoints will be deprecated in favor of new endpoints. Both versions will be available in 0.25 but the deprecated endpoints will be removed in a subsequent release.
For master endpoints:
For slave endpoints:
For both master and slave:
NOTE The C++/Java/Python scheduler bindings have been updated. In particular, the driver can make a suppressOffers() call to stop receiving offers (until reviveOffers() is called).
In order to upgrade a running cluster:
NOTE Support for live upgrading a driver based scheduler to HTTP based (experimental) scheduler has been added.
NOTE Master now publishes its information in ZooKeeper in JSON (instead of protobuf). Make sure schedulers are linked against >= 0.23.0 libmesos before upgrading the master.
In order to upgrade a running cluster:
NOTE The ‘stats.json’ endpoints for masters and slaves have been removed. Please use the ‘metrics/snapshot’ endpoints instead.
NOTE The ‘/master/shutdown’ endpoint is deprecated in favor of the new ‘/master/teardown’ endpoint.
NOTE In order to enable decorator modules to remove metadata (environment variables or labels), we changed the meaning of the return value for decorator hooks in Mesos 0.23.0. Please refer to the modules documentation for more details.
NOTE Slave ping timeouts are now configurable on the master via --slave_ping_timeout
and --max_slave_ping_timeouts
. Slaves should be upgraded to 0.23.x before changing these flags.
NOTE A new scheduler driver API, acceptOffers
, has been introduced. This is a more general version of the launchTasks
API, which allows the scheduler to accept an offer and specify a list of operations (Offer.Operation) to perform using the resources in the offer. Currently, the supported operations include LAUNCH (launching tasks), RESERVE (making dynamic reservations), UNRESERVE (releasing dynamic reservations), CREATE (creating persistent volumes) and DESTROY (releasing persistent volumes). Similar to the launchTasks
API, any unused resources will be considered declined, and the specified filters will be applied on all unused resources.
NOTE The Resource protobuf has been extended to include more metadata for supporting persistence (DiskInfo), dynamic reservations (ReservationInfo) and oversubscription (RevocableInfo). You must not combine two Resource objects if they have different metadata.
In order to upgrade a running cluster:
NOTE Slave checkpoint flag has been removed as it will be enabled for all slaves. Frameworks must still enable checkpointing during registration to take advantage of checkpointing their tasks.
NOTE The stats.json endpoints for masters and slaves have been deprecated. Please refer to the metrics/snapshot endpoint.
NOTE The C++/Java/Python scheduler bindings have been updated. In particular, the driver can be constructed with an additional argument that specifies whether to use implicit driver acknowledgements. In statusUpdate
, the TaskStatus
now includes a UUID to make explicit acknowledgements possible.
NOTE: The Authentication API has changed slightly in this release to support additional authentication mechanisms. The change from ‘string’ to ‘bytes’ for AuthenticationStartMessage.data has no impact on C++ or the over-the-wire representation, so it only impacts pure language bindings for languages like Java and Python that use different types for UTF-8 strings vs. byte arrays.
message AuthenticationStartMessage { required string mechanism = 1; optional bytes data = 2; }
NOTE All Mesos arguments can now be passed using file:// to read them out of a file (either an absolute or relative path). The --credentials, --whitelist, and any flags that expect JSON backed arguments (such as --modules) behave as before, although support for just passing an absolute path for any JSON flags rather than file:// has been deprecated and will produce a warning (and the absolute path behavior will be removed in a future release).
In order to upgrade a running cluster:
MesosSchedulerDriverImpl
with Credentials
, your code must be updated to pass the implicitAcknowledgements
argument before Credentials
. You may run a 0.21.0 Python scheduler against a 0.22.0 master, and vice versa.NOTE Disabling slave checkpointing has been deprecated; the slave --checkpoint flag has been deprecated and will be removed in a future release.
In order to upgrade a running cluster:
NOTE: The Mesos API has been changed slightly in this release. The CommandInfo has been changed (see below), which makes launching a command more flexible. The ‘value’ field has been changed from required to optional. However, it will not cause any issue during the upgrade (since the existing schedulers always set this field).
message CommandInfo { ... // There are two ways to specify the command: // 1) If 'shell == true', the command will be launched via shell // (i.e., /bin/sh -c 'value'). The 'value' specified will be // treated as the shell command. The 'arguments' will be ignored. // 2) If 'shell == false', the command will be launched by passing // arguments to an executable. The 'value' specified will be // treated as the filename of the executable. The 'arguments' // will be treated as the arguments to the executable. This is // similar to how POSIX exec families launch processes (i.e., // execlp(value, arguments(0), arguments(1), ...)). optional bool shell = 6 [default = true]; optional string value = 3; repeated string arguments = 7; ... }
NOTE: The Python bindings are also changing in this release. There are now sub-modules which allow you to use either the interfaces and/or the native driver.
import mesos.native
for the native driversimport mesos.interface
for the stub implementations and protobufsTo ensure a smooth upgrade, we recommend to upgrade your python framework and executor first. You will be able to either import using the new configuration or the old. Replace the existing imports with something like the following:
try: from mesos.native import MesosExecutorDriver, MesosSchedulerDriver from mesos.interface import Executor, Scheduler from mesos.interface import mesos_pb2 except ImportError: from mesos import Executor, MesosExecutorDriver, MesosSchedulerDriver, Scheduler import mesos_pb2
NOTE: If you're using a pure language binding, please ensure that it sends status update acknowledgements through the master before upgrading.
In order to upgrade a running cluster:
NOTE: There are new required flags on the master (--work_dir
and --quorum
) to support the Registrar feature, which adds replicated state on the masters.
NOTE: No required upgrade ordering across components.
In order to upgrade a running cluster:
In order to upgrade a running cluster:
NOTE: This upgrade requires a system reboot for slaves that use Linux cgroups for isolation.
In order to upgrade a running cluster:
In order to upgrade a running cluster:
In order to upgrade a running cluster:
reconcileTasks
driver method.MesosSchedulerDriver
constructor that takes Credential
to authenticate.In order to upgrade a running cluster:
In order to upgrade a running cluster:
In order to upgrade a running cluster:
If you are a framework developer, you will want to examine the new ‘source’ field in the ExecutorInfo protobuf. This will allow you to take further advantage of the resource monitoring.