The task manager is a generic system for running background tasks.
Documentation: https://www.elastic.co/guide/en/kibana/current/task-manager-production-considerations.html
It supports:
- Single-run and recurring tasks
- Scheduling tasks to run after a specified datetime
- Basic retry logic
- Recovery of stalled tasks / timeouts
- Tracking task state across multiple runs
- Configuring the run-parameters for specific tasks
- Basic coordination to prevent the same task instance from running on more than one Kibana system at a time
At a high-level, the task manager works like this:
- Every
{poll_interval}milliseconds, check the{index}for any tasks that need to be run:runAtis pastattemptsis less than the configured threshold
- Attempt to claim the task by using optimistic concurrency to set:
- status to
running startedAtto nowretryAtto next time task should retry if it times out and is still inrunningstatus
- status to
- Execute the task, if the previous claim succeeded
- If the task fails, increment the
attemptscount and reschedule it - If the task succeeds:
- If it is recurring, store the result of the run, and reschedule
- If it is not recurring, remove it from the index
Each task manager instance runs tasks in a pool which ensures that at most N tasks are run at a time, where N is configurable. This prevents the system from running too many tasks at once in resource constrained environments. In addition to this, each individual task type definition can have numWorkers specified, which tells the system how many workers are consumed by a single running instance of a task. This effectively limits how many tasks of a given type can be run at once.
For example, we may have a system with a max_workers of 10, but a super expensive task (such as reporting) which specifies a numWorkers of 10. In this case, reporting tasks will run one at a time.
If a task specifies a higher numWorkers than the system supports, the system's max_workers setting will be substituted for it.
The task_manager can be configured via taskManager config options (e.g. taskManager.maxAttempts):
max_attempts- The maximum number of times a task will be attempted before being abandoned as failedpoll_interval- How often the background worker should check the task_manager index for more workmax_poll_inactivity_cycles- How many poll intervals is work allowed to block polling for before it's timed out. This does not include task execution, as task execution does not block the polling, but rather includes work needed to manage Task Manager's state.index- deprecated The name of the index that the task_manager will use. This is deprecated, and will be removed starting in 8.0max_workers- The maximum number of tasks a Kibana will run concurrently (defaults to 10)version_conflict_threshold- The threshold percentage for workers experiencing version conflicts for shifting the polling intervalcredentials- Encrypted user credentials. All tasks will run in the security context of this user. See this issue for a discussion on task scheduler security.override_num_workers: An object oftaskType: numberthat overrides thenum_workersfor tasks- For example:
task_manager.override_num_workers.reporting: 2would override the number of workers occupied by tasks of typereporting - This allows sysadmins to tweak the operational performance of Kibana, allowing more or fewer tasks of a specific type to run simultaneously
- For example:
monitored_aggregated_stats_refresh_rate- Dictates how often we refresh the "Cold" metrics. Learn More: ./MONITORINGmonitored_stats_running_average_window- Dictates the size of the window used to calculate the running average of various "Hot" stats. Learn More: ./MONITORINGmonitored_stats_required_freshness- Dictates the required freshness of critical "Hot" stats. Learn More: ./MONITORINGmonitored_task_execution_thresholds- Dictates the threshold of failed task executions. Learn More: ./MONITORINGunsafe.exclude_task_types- A list of task types to exclude from running. Supports wildcard usage, such asnamespace:*. This configuration is experimental, unsupported, and can only be used for temporary debugging purposes because it causes Kibana to behave in unexpected ways.
Plugins define tasks by calling the registerTaskDefinitions method on the server.plugins.task_manager object.
A sample task can be found in the x-pack/test/plugin_api_integration/plugins/task_manager folder.
export class Plugin {
constructor() {
}
public setup(core: CoreSetup, plugins: { taskManager }) {
taskManager.registerTaskDefinitions({
// clusterMonitoring is the task type, and must be unique across the entire system
clusterMonitoring: {
// Human friendly name, used to represent this task in logs, UI, etc
title: 'Human friendly name',
// Optional, human-friendly, more detailed description
description: 'Amazing!!',
// Optional, how long, in minutes or seconds, the system should wait before
// a running instance of this task is considered to be timed out.
// This defaults to 5 minutes.
timeout: '5m',
// Optional, how many attempts before marking task as failed.
// This defaults to what is configured at the task manager level.
maxAttempts: 5,
// The maximum number tasks of this type that can be run concurrently per Kibana instance.
// Setting this value will force Task Manager to poll for this task type seperatly from other task types which
// can add significant load to the ES cluster, so please use this configuration only when absolutly necesery.
maxConcurrency: 1,
// The createTaskRunner function / method returns an object that is responsible for
// performing the work of the task. context: { taskInstance }, is documented below.
createTaskRunner(context) {
return {
// Perform the work of the task. The return value should fit the TaskResult interface, documented
// below. Invalid return values will result in a logged warning.
async run() {
// Do some work
// Conditionally send some alerts
// Return some result or other...
},
// Optional, will be called if a running instance of this task times out, allowing the task
// to attempt to clean itself up.
async cancel() {
// Do whatever is required to cancel this task, such as killing any spawned processes
},
};
},
},
});
}
public start(core: CoreStart, plugins: { taskManager }) {
}
}When Kibana attempts to claim and run a task instance, it looks its definition up, and executes its createTaskRunner's method, passing it a run context which looks like this:
{
// The data associated with this instance of the task, with two properties being most notable:
//
// params:
// An object, specific to this task instance, used by the
// task to determine exactly what work should be performed.
// e.g. a cluster-monitoring task might have a `clusterName`
// property in here, but a movie-monitoring task might have
// a `directorName` property.
//
// state:
// The state returned from the previous run of this task instance.
// If this task instance has never succesfully run, this will
// be an empty object: {}
taskInstance,
}The task runner's run method is expected to return a promise that resolves to undefined or to an object that looks like the following:
| Property | Description | Type |
|---|---|---|
| runAt | Optional. If specified, this is used as the tasks' next runAt, overriding the default system scheduler. |
Date ISO String |
| schedule | Optional. If specified, this is used as the tasks' new recurring schedule, overriding the default system scheduler and any existing schedule. | { interval: string } |
| error | Optional, an error object, logged out as a warning. The pressence of this property indicates that the task did not succeed. | Error |
| state | Optional, this will be passed into the next run of the task, if this is a recurring task. | Record<string, unknown> |
{
// Optional, if specified, this is used as the tasks' nextRun, overriding
// the default system scheduler.
runAt: "2020-07-24T17:34:35.272Z",
error: { message: 'Hrumph!' },
state: {
anything: 'goes here',
},
}{
schedule: { interval: '30s' },
state: {
anything: 'goes here',
},
}Other return values will result in a warning, but the system should continue to work.
If a task runner throws an error, task manager will try to rerun the task shortly after (up to the task definition's maxAttempts).
Normal tasks will wait a default amount of 5m before trying again and every subsequent attempt will add an additonal 5m cool off period to avoid a stampeding herd of failed tasks from storming Elasticsearch.
Recurring tasks will also get retried, but instead of using the 5m interval for the retry, they will be retried on their next scheduled run.
If you wish to purposfully fail a task, you can throw an error of any kind and the retry logic will apply.
If, on the other hand, you wish not only to fail the task, but you'd also like to indicate the Task Manager that it shouldn't retry the task, you can throw an Unrecoverable Error, using the throwUnrecoverableError helper function.
For example:
taskManager.registerTaskDefinitions({
myTask: {
/// ...
createTaskRunner(context) {
return {
async run() {
const result = ... // Do some work
if(!result) {
// No point retrying?
throwUnrecoverableError(new Error("No point retrying, this is unrecoverable"));
}
return result;
}
};
},
},
});The task_manager module will store scheduled task instances in an index. This allows for recovery of failed tasks, coordination across Kibana clusters, persistence across Kibana reboots, etc.
The data stored for a task instance looks something like this:
{
// The type of task that will run this instance.
taskType: 'clusterMonitoring',
// The next time this task instance should run. It is not guaranteed
// to run at this time, but it is guaranteed not to run earlier than
// this.
runAt: "2020-07-24T17:34:35.272Z",
// Indicates that this is a recurring task. We support interval syntax
// with days such as '1d', hours '3h', minutes such as `5m`, seconds `10s`.
schedule: { interval: '5m' },
// How many times this task has been unsuccesfully attempted,
// this will be reset to 0 if the task ever succesfully completes.
// This is incremented if a task fails or times out.
attempts: 0,
// Currently, this is either idle | claiming | running | failed. It is used to
// coordinate which Kibana instance owns / is running a specific
// task instance.
// idle: Task Instance isn't being worked on
// claiming: A Kibana instance has claimed ownership but hasn't started running
// the Task Instance yet
// running: A Kibana instance has began working on the Task Instance
// failed: The last run of the Task Instance failed, waiting to retry
status: 'idle',
// The params specific to this task instance, which will be
// passed to the task when it runs, and will be used by the
// task to determine exactly what work should be performed.
// This is a JSON blob, and will be different per task type.
// e.g. a cluster-monitoring task might have a `clusterName`
// property in here, but a movie-monitoring task might have
// a `directorName` property.
params: '{ "task": "specific stuff here" }',
// The result of the previous run of this task instance. This
// will be passed to the next run of the task, along with the
// params, and could be used by a task to do special logic If
// the task state changes (e.g. from green to red, or foo to bar)
// If there was no previous run (e.g. the instance has never succesfully
// completed, this will be an empty object.). This is a JSON blob,
// and will be different per task type.
state: '{ "status": "green" }',
// An extension point for 3rd parties to build in security features on
// top of the task manager. For example, this might be the token of the user
// who scheduled this task.
userContext: 'the token of the user who scheduled this task',
// An extension point for 3rd parties to build in security features on
// top of the task manager, and is expected to be the id of the user, if any,
// that scheduled this task.
user: '23lk3l42',
// An application-specific designation, allowing different Kibana
// plugins / apps to query for only those tasks they care about.
scope: ['alerting'],
// The Kibana UUID of the Kibana instance who last claimed ownership for running this task.
ownerId: '123e4567-e89b-12d3-a456-426655440000'
}The task manager mixin exposes a taskManager object on the Kibana server which plugins can use to manage scheduled tasks. Each method takes an optional scope argument and ensures that only tasks with the specified scope(s) will be affected.
Interaction with the TaskManager Plugin is done via the Kibana Platform Plugin system.
When developing your Plugin, you're asked to define a setup method and a start method.
These methods are handed Kibana's Plugin APIs for these two stages, which means you'll have access to the following apis in these two stages:
The Setup Plugin api includes methods which configure Task Manager to support your Plugin's requirements, such as defining custom Middleware and Task Definitions.
{
addMiddleware: (middleware: Middleware) => {
// ...
},
registerTaskDefinitions: (taskDefinitions: TaskDictionary<TaskDefinition>) => {
// ...
},
}The Start Plugin api allow you to use Task Manager to facilitate your Plugin's behaviour, such as scheduling tasks.
{
fetch: (opts: FetchOpts) => {
// ...
},
remove: (id: string) => {
// ...
},
get: (id: string) => {
// ...
},
schedule: (taskInstance: TaskInstanceWithDeprecatedFields, options?: any) => {
// ...
},
runNow: (taskId: string) => {
// ...
},
ensureScheduled: (taskInstance: TaskInstanceWithId, options?: any) => {
// ...
},
}Using schedule you can instruct TaskManger to schedule an instance of a TaskType at some point in the future.
export class Plugin {
constructor() {
}
public setup(core: CoreSetup, plugins: { taskManager }) {
}
public start(core: CoreStart, plugins: { taskManager }) {
// Schedules a task. All properties are as documented in the previous
// storage section, except that here, params is an object, not a JSON
// string.
const task = await taskManager.schedule({
taskType,
runAt,
schedule,
params,
scope: ['my-fanci-app'],
});
// Removes the specified task
await taskManager.remove(task.id);
// Fetches tasks, supports pagination, via the search-after API:
// https://www.elastic.co/guide/en/elasticsearch/reference/current/search-request-search-after.html
// If scope is not specified, all tasks are returned, otherwise only tasks
// with the given scope are returned.
const results = await taskManager.find({ scope: 'my-fanci-app', searchAfter: ['ids'] });
}
}results then look something like this:
{
"searchAfter": ["233322"],
// Tasks is an array of task instances
"tasks": [{
"id": "3242342",
"taskType": "reporting",
// etc
}]
}When using the schedule api to schedule a Task you can provide a hard coded id on the Task. This tells TaskManager to use this id to identify the Task Instance rather than generate an id on its own.
The danger is that in such a situation, a Task with that same id might already have been scheduled at some earlier point, and this would result in an error. In some cases, this is the expected behavior, but often you only care about ensuring the task has been scheduled and don't need it to be scheduled a fresh.
To achieve this you should use the ensureScheduling api which has the exact same behavior as schedule, except it allows the scheduling of a Task with an id that's already in assigned to another Task and it will assume that the existing Task is the one you wished to schedule, treating this as a successful operation.
Using runNow you can instruct TaskManger to run an existing task on-demand, without waiting for its scheduled time to be reached.
export class Plugin {
constructor() {
}
public setup(core: CoreSetup, plugins: { taskManager }) {
}
public start(core: CoreStart, plugins: { taskManager }) {
try {
const taskRunResult = await taskManager.runNow('91760f10-ba42-de9799');
// If no error is thrown, the task has completed successfully.
} catch(err: Error) {
// If running the task has failed, we throw an error with an appropriate message.
// For example, if the requested task doesnt exist: `Error: failed to run task "91760f10-ba42-de9799" as it does not exist`
// Or if, for example, the task is already running: `Error: failed to run task "91760f10-ba42-de9799" as it is currently running`
}
}
}More custom access to the tasks can be done directly via Elasticsearch, though that won't be officially supported, as we can change the document structure at any time.
The task manager exposes a middleware layer that allows modifying tasks before they are scheduled / persisted to the task manager index, and modifying tasks / the run context before a task is run.
For example:
export class Plugin {
constructor() {
}
public setup(core: CoreSetup, plugins: { taskManager }) {
taskManager.addMiddleware({
async beforeSave({ taskInstance, ...opts }) {
console.log(`About to save a task of type ${taskInstance.taskType}`);
return {
...opts,
taskInstance: {
...taskInstance,
params: {
...taskInstance.params,
example: 'Added to params!',
},
},
};
},
async beforeRun({ taskInstance, ...opts }) {
console.log(`About to run ${taskInstance.taskType} ${taskInstance.id}`);
const { example, ...taskWithoutExampleProp } = taskInstance;
return {
...opts,
taskInstance: taskWithoutExampleProp,
};
},
});
}
public start(core: CoreStart, plugins: { taskManager }) {
}
}TaskManager used to work in a pull model, but it now needs to support both push and pull, so it has been remodeled internally to support a single push model.
Task Manager's push mechanism is driven by the following operations:
- A polling interval has been reached.
- A new Task is scheduled.
- A Task is run using
runNow.
The polling interval is straight forward: TaskPoller is configured to emit an event at a fixed interval. That said, if there are no workers available, we want to ignore these events, so we'll throttle the interval on worker availability.
Whenever a user uses the schedule api to schedule a new Task, we want to trigger an early polling in order to respond to the newly scheduled task as soon as possible, but this too we only wish to do if there are available workers, so we can throttle this too.
When a runNow call is made we need to force a poll as the user will now be waiting on the result of the runNow call, but
there is a complexity here- we don't want to force polling (as there might not be any worker capacity and it's possible that a polling cycle is already running), but we also can't throttle, as we can't afford to "drop" these requests, so we'll have to buffer these.
We now want to respond to all three of these push events, but we still need to balance against our worker capacity, so if there are too many requests buffered, we only want to take as many requests as we have capacity to handle.
Luckily, Polling Interval and Task Scheduled simply denote a request to "poll for work as soon as possible", unlike Run Task Now which also means "poll for these specific tasks", so our worker capacity only needs to be applied to Run Task Now.
We achieve this model by buffering requests into a queue using a Set (which removes duplicated). As we don't want an unbounded queue in our system, we have limited the size of this queue (configurable by the xpack.task_manager.request_capacity config, defaulting to 1,000 requests) which forces us to throw an error once this cap is reachedand to all subsequent calls to runNow until the queue drain bellow the cap.
Our current model, then, is this:
Polling Interval --> filter(availableWorkers > 0) - mapTo([]) -------\\
Task Scheduled --> filter(availableWorkers > 0) - mapTo([]) --------||==>Set([]+[]+[`1`,`2`]) ==> work([`1`,`2`])
Run Task `1` Now --\ //
----> buffer(availableWorkers > 0) -- [`1`,`2`] -//
Run Task `2` Now --/
There is only a rudimentary mechanism for coordinating tasks and handling expired tasks. Tasks are considered expired if their runAt has arrived, and their status is still 'running'.
There is no task history. Each run overwrites the previous run's state. One-time tasks are removed from the index upon completion.
The task manager's public API is create / delete / list. Updates aren't directly supported, and listing should be scoped so that users only see their own tasks.
-
Unit tests:
Documentation: https://www.elastic.co/guide/en/kibana/current/development-tests.html#_unit_testing
yarn test:jest x-pack/plugins/task_manager --watch -
Integration tests:
node scripts/functional_tests_server.js --config x-pack/test/plugin_api_integration/config.ts node scripts/functional_test_runner --config x-pack/test/plugin_api_integration/config.ts
Task Manager exposes runtime statistics which enable basic observability into its inner workings and makes it possible to monitor the system from external services.
Public Documentation: https://www.elastic.co/guide/en/kibana/master/task-manager-health-monitoring.html Developer Documentation: ./MONITORING