Implement Sahara cluster verification checks

Implement Sahara cluster verification checks

https://blueprints.launchpad.net/sahara/+spec/cluster-verification

Now we don’t have any check to take a look at cluster processes status through Sahara interface. Our plans is to implement cluster verifications and ability to re-trigger these verifications for the particular cluster.

Problem description

Sahara doesn’t have any check for cluster processes health monitoring. Cluster can be broken or unavailable but Sahara will still think that in ACTIVE status. This may result in end-users losses and so on.

Proposed change

First of all let’s make several important definitions in here. A cluster health check is a perform a limited functionality check of a cluster. For example, instances accessibility, writing some data to HDFS and so on. Each health check will be implemented as a class object, implementing several important methods from an abstract base class(abc):

class ExampleHealthCheck(object):
    def __init__(self, cluster, *args, **kwargs):
        self.cluster = cluster
        # other stuff

    @abc.abstractmethod
    def available(self):
        # method will verify

    @abc.abstractmethod
    def check(self):
        # actual health will in here

    def execute(self):
        # based in availability of the check and results
        # of check will write correct data into database

The expected behaviour of check method of a health check is return some important data in case when everything is ok and in case of errors to raise an exception with detailed info with failures reasons. Let’s describe important statuses of the health checks.

  • GREEN status: cluster in healthy and health check passed correctly. In description in such case we can have the following info: HDFS available for writing data or All datanodes are active and available.
  • YELLOW status: YellowHealthError will be raised as result of the operation, it means that something is probably wrong with cluster. By the way cluster is still operable and can be used for running jobs. For example, in exception message we will see the following information: 2 out of 10 datanodes are not working.
  • RED status: RedHealthError will be raised in such case, it means that something is definitely wrong we the cluster we can’t guarantee that cluster is still able to perform perations on it. For example, Amount of active datanodes is less then replication is possible messange in such case.
  • CHECKING state, health check is still running or was just created.

A cluster verification is a combination of the several cluster health checks. Cluster verification will indicate the current status for the cluster: GREEN, YELLOW or RED. We will store the latest verification for the cluster in database. Also we will send results of verifications to Ceilometer, to view progress of cluster health.

Also there is an idea of running several jobs as part of several health checks, but it will be too harmful for the cluster health and probably will be done later.

Also we can introduce periodic task for refreshing health status of the cluster.

So, several additional options in sahara.conf should be added in new section cluster_health_verification:

  • enable_health_verification by default will be True, will allow disabling periodic cluster verifications
  • verification_period variable to define period between two consecutive health verifications in periodic tasks. By default I would suggest to run verification once in 10 minutes.

Proposed checks

This section is going to describe basic checks of functionalities of clusters. Several checks will affect almost all plugins, few checks will be specific only for the one plugin.

Basic checks:

There are several basic checks for all possible clusters in here.

  • Check to verify all instances access. If some instances are not accessible we have RED state.
  • Check that all volumes mounted. If some volume is not mounted, we will have YELLOW state.

HDFS checks:

  • Check that will verify that namenode is working. Of course, RED state in bad case. Actually this will affects only vanilla plugin clusters and clusters deployed without HA mode in case of CDH and Ambari plugins.
  • Check that will verify amount of living datanodes. We will have GREEN status only when all datanodes are active, RED in case when amount of living datanodes are less then dfs.replicaton, and YELLOW in all other cases.
  • Check to verify ability of cluster to write some data to HDFS. We will have RED status if something is failed.
  • Amount of free space in HDFS. We will compare this value with reserved memory in HDFS, and if amount of free space is less then provided value, check will be supposed to be failed. If check is not passed, we will have YELLOW state we will advice to scale cluster up with some extra datanodes (or clean some data). I think, we should not have any additional configuration options in here, just because this check will never report RED state.

HA checks:

  • YELLOW state in case when at least one stand-by service is working; and RED otherwise. Affects YARN and HDFS both.

YARN checks:

  • Resourcemanger is active. Obviously, RED state if something is wrong.
  • Amount of active nodemanagers. YELLOW state if something is not available, and RED if amout of live nodemanagers are less then 50%.

Kafka check:

  • Check that kafka is operable: create example topic, put several messages in topic, consuming messages. RED state in case of something is wrong.

CDH plugin check:

This section is going to describe specific checks for CDH plugin. For this checks we will need to extend current sahara’s implementation of cm_api tool. There is an API methods to get current health of the cluster. There are few examples of responses for yarn service.

There is the bad case example:

"yarn01": {
  "checks": [
    {
      "name": "YARN_JOBHISTORY_HEALTH",
      "summary": "GOOD"
    },
    {
      "name": "YARN_NODE_MANAGERS_HEALTHY",
      "summary": "CONCERNING"
    },
    {
      "name": "YARN_RESOURCEMANAGERS_HEALTH",
      "summary": "BAD"
    }
  ],
  "summary": "BAD"
}

and good case example:

"yarn01": {
  "checks": [
    {
      "name": "YARN_JOBHISTORY_HEALTH",
      "summary": "GOOD"
    },
    {
      "name": "YARN_NODE_MANAGERS_HEALTHY",
      "summary": "GOOD"
    },
    {
      "name": "YARN_RESOURCEMANAGERS_HEALTH",
      "summary": "GOOD"
    }
  ],
  "summary": "GOOD"
}

Based on responses above we will calculate health of the cluster. Also possible states which Cloudera can return through API are DISABLED when service was stopped and CONCERNING if something is going to be bad soon. In this health check sahara’s statuses will be calculated based on the following table:

+--------------+--------------------------------+
| Sahara state |   Cloudera state               |
+--------------+--------------------------------+
| GREEN        | All services GOOD              |
+--------------+--------------------------------+
| YELLOW       | At least 1 service CONCERNING  |
+--------------+--------------------------------+
| RED          | At least 1 service BAD/DISABLED|
+--------------+--------------------------------+

Some additional information about Cloudera health checks are in here: [0]

Ambari plugin:

Current HDP 2.0.6 will support only basic verifications. The main focus in here is to implement additional checks for the Ambari plugin. There are several ideas of checks in Ambari plugin:

  • Ambari alerts verification. Ambari plugin have several alerts if something is wrong with current state of the cluster. We can get alerts through Ambari API. If we have at least one alert in here it’s proposed to use YELLOW status for the verification, and otherwise we will use GREEN status for that.
  • Ambari service checks verification. Ambari plugin have a bunch of services checks in here, which can be re-triggered by user through the Ambari API. These checks are well described in [1]. If at least one check failed, we will use RED status for that sutiation, otherwise it’s nice to use GREEN.

Alternatives

All health checks can be disabled by the option.

Data model impact

Graphical description of data model impact:

+----------------------------+    +-------------------------------+
|       verifications        |    |           health_checks       |
+----------------------------+    +-----------------+-------------+
|    id      | Primary Key   |    | id              | Primary Key |
+------------+---------------+    +-----------------+-------------+
| cluster_id | Foreign Key   |  +-| verification_id | Foreign Key |
+----------------------------+  | +-----------------+-------------+
| created_at |               |  | | created_at      |             |
+------------+---------------+  | +-----------------+-------------+
| updated_at |               |  | | updated_at      |             |
+------------+---------------+  | +-----------------+-------------+
| checks     |               | <+ | status          |             |
+------------+---------------+    +-----------------+-------------+
| status     |               |    | description     |             |
+------------+---------------+    +-----------------+-------------+
                                  | name            |             |
                                  +-----------------+-------------+

We will have two additional tables where we will store verifications and health checks.

First table with of verifications will have following columns id, cluster_id (foreign key), created_at, updated_at.

Also will be added new table to store health check results. This table will have the following columns: id, verification_id, description, status, created_at and updated_at.

We will have cascade relationship (checks) between cluster verifications and cluster health checks to get correct access from health check to cluster verification and vice versa. Also same relationship will be between cluster and verification for same purpose.

Also to aggregation results of latest verification and disabling/enabling verifications for particular cluster will be added the new column to cluster model: verifications_status. We will not use status for that purpose just to keep these two variables separately (we already using status in many places in sahara).

For example of verifications:

  1. One health check is still running:
"cluster_verification": {
  "id": "1",
  "cluster_id": "1111",
  "created_at": "2013-10-09 12:37:19.295701",
  "updated_at": "2013-10-09 12:37:19.295701",
  "status": "CHECKING",
  "checks": [
    {
      "id": "123",
      "created_at": "2013-10-09 12:37:19.295701",
      "updated_at": "2013-10-09 12:37:19.295701",
      "status": "GREEN",
      "description": "some description",
      "name": "same_name"
    },
    {
      "id": "221",
      "created_at": "2013-10-09 12:37:19.295701",
      "updated_at": "2013-10-09 12:37:19.295701",
      "status": "CHECKING",
      "description": "some description",
      "name": "same_name"
    },
  ]
}
  1. All health checks are completed but one was failed:
"cluster_verification": {
  "id": "2",
  "cluster_id": "1112",
  "created_at": "2013-10-09 12:37:19.295701",
  "updated_at": "2013-10-09 12:37:30.295701",
  "STATUS": "RED",
  "checks": [
    {
      ..
      "status": "RED",
      "description": "Resourcemanager is down",
      ..
    },
    {
      ..
      "status": "GREEN",
      "description": "HDFS is healthy",
    }
  ]
}

REST API impact

Mechanism of receiving results of cluster verifications will be quite simple. We will just use usual GET method for clusters.

So, the main API method will be the following: GET <tenant_id>/clusters/<cluster_id>. In such case, we will return detailed info of the cluster with verifications.

Example of response:

{
  "status": "Active",
  "id": "1111",
  "cluster_template_id": "5a9a09a3-9349-43bd-9058-16c401fad2d5",
  "name": "sample",
  "verifications_status": "RUNNING",
  ..
  "verification": {
    "id": "1",
    "cluster_id": "1111",
    "created_at": "2013-10-09 12:37:19.295701",
    "updated_at": "2013-10-09 12:37:19.295701",
    "checks": [
      {
        "id": "123",
        "created_at": "2013-10-09 12:37:19.295701",
        "updated_at": "2013-10-09 12:37:19.295701",
        "status": "GREEN",
        "description": "some description",
        "name": "same_name"
      },
      {
        "id": "221",
        "created_at": "2013-10-09 12:37:19.295701",
        "updated_at": "2013-10-09 12:37:19.295701",
        "status": "CHECKING",
        "description": "some description",
        "name": "same_name"
      },
    ]
  }
}

For re-triggering to cluster verification, some additional behaviour should be added to the following API method:

PATCH <tenant_id>/clusters/<cluster_id>

If the following data will be provided to this API method we will re-trigger verification:

{
  'verification': {
    'status': 'START'
  }
}

Start will be reject when verifications disabled for the cluster or when verification is running on the cluster.

Also we can disable verification for particular cluster to avoid unneeded noisy verifications until health issues are fixed by the following request data:

{
  'verification': {
    'status': 'DISABLE'
  }
}

And enable in case we need to enable health checks again. If user is trying to disable verification only future verifications will be disabled, so health checks still will be running.

If something additional will be added to this data we will mark request as invalid. Also we will implement new validation methods to deny verifications on cluster which already have one verification running.

Other end user impact

Need to implement requests for run checks via python-saharaclient and get their results.

Deployer impact

None.

Developer impact

None.

Sahara-image-elements impact

None.

Sahara-dashboard / Horizon impact

Dashboard impact is need to add new tab in cluster details with results of verifications.

Implementation

Assignee(s)

Primary assignee:
vgridnev
Other contributors:
apavlov-n, esikachev

Work Items

  1. Implement basic skeleton for verifications (with base checks)
  2. Python-saharaclient support addition
  3. CLI support should be implemented
  4. Implement tab with verification results to Horizon
  5. Need to add new WADL docs with new api-method
  6. All others checks should be implemented
  7. Should be added support to scenario framework to allow re-triggering.
  8. Implement sending history to Ceilometer.

Dependencies

None

Testing

Feature will be covered by the unit tests, and manually. New test commit (not for merging) will be added to show that all verifications are passed (since we are at the middle of moving scenario framework).

Documentation Impact

Documentaion should updated with additional information of How to repair issues described in the health check results.

Creative Commons Attribution 3.0 License

Except where otherwise noted, this document is licensed under Creative Commons Attribution 3.0 License. See all OpenStack Legal Documents.

sahara-specs 0.0.1.dev332