Skip to content

Latest commit

 

History

History
738 lines (567 loc) · 22.3 KB

fix-common-cluster-issues.asciidoc

File metadata and controls

738 lines (567 loc) · 22.3 KB

Fix common cluster issues

This guide describes how to fix common errors and problems with {es} clusters.

Error: disk usage exceeded flood-stage watermark, index has read-only-allow-delete block

This error indicates a data node is critically low on disk space and has reached the flood-stage disk usage watermark. To prevent a full disk, when a node reaches this watermark, {es} blocks writes to any index with a shard on the node. If the block affects related system indices, {kib} and other {stack} features may become unavailable.

{es} will automatically remove the write block when the affected node’s disk usage goes below the high disk watermark. To achieve this, {es} automatically moves some of the affected node’s shards to other nodes in the same data tier.

To verify that shards are moving off the affected node, use the cat shards API.

GET _cat/shards?v=true

If shards remain on the node, use the cluster allocation explanation API to get an explanation for their allocation status.

GET _cluster/allocation/explain
{
  "index": "my-index",
  "shard": 0,
  "primary": false,
  "current_node": "my-node"
}

To immediately restore write operations, you can temporarily increase the disk watermarks and remove the write block.

PUT _cluster/settings
{
  "persistent": {
    "cluster.routing.allocation.disk.watermark.low": "90%",
    "cluster.routing.allocation.disk.watermark.high": "95%",
    "cluster.routing.allocation.disk.watermark.flood_stage": "97%"
  }
}

PUT */_settings?expand_wildcards=all
{
  "index.blocks.read_only_allow_delete": null
}

As a long-term solution, we recommend you add nodes to the affected data tiers or upgrade existing nodes to increase disk space. To free up additional disk space, you can delete unneeded indices using the delete index API.

DELETE my-index

When a long-term solution is in place, reset or reconfigure the disk watermarks.

PUT _cluster/settings
{
  "persistent": {
    "cluster.routing.allocation.disk.watermark.low": null,
    "cluster.routing.allocation.disk.watermark.high": null,
    "cluster.routing.allocation.disk.watermark.flood_stage": null
  }
}

Circuit breaker errors

{es} uses circuit breakers to prevent nodes from running out of JVM heap memory. If Elasticsearch estimates an operation would exceed a circuit breaker, it stops the operation and returns an error.

By default, the parent circuit breaker triggers at 95% JVM memory usage. To prevent errors, we recommend taking steps to reduce memory pressure if usage consistently exceeds 85%.

Diagnose circuit breaker errors

Error messages

If a request triggers a circuit breaker, {es} returns an error with a 429 HTTP status code.

{
  'error': {
    'type': 'circuit_breaking_exception',
    'reason': '[parent] Data too large, data for [<http_request>] would be [123848638/118.1mb], which is larger than the limit of [123273216/117.5mb], real usage: [120182112/114.6mb], new bytes reserved: [3666526/3.4mb]',
    'bytes_wanted': 123848638,
    'bytes_limit': 123273216,
    'durability': 'TRANSIENT'
  },
  'status': 429
}

{es} also writes circuit breaker errors to elasticsearch.log. This is helpful when automated processes, such as allocation, trigger a circuit breaker.

Caused by: org.elasticsearch.common.breaker.CircuitBreakingException: [parent] Data too large, data for [<transport_request>] would be [num/numGB], which is larger than the limit of [num/numGB], usages [request=0/0b, fielddata=num/numKB, in_flight_requests=num/numGB, accounting=num/numGB]

Check JVM memory usage

If you’ve enabled Stack Monitoring, you can view JVM memory usage in {kib}. In the main menu, click Stack Monitoring. On the Stack Monitoring Overview page, click Nodes. The JVM Heap column lists the current memory usage for each node.

You can also use the cat nodes API to get the current heap.percent for each node.

GET _cat/nodes?v=true&h=name,node*,heap*

To get the JVM memory usage for each circuit breaker, use the node stats API.

GET _nodes/stats/breaker

Prevent circuit breaker errors

Reduce JVM memory pressure

High JVM memory pressure often causes circuit breaker errors. See High JVM memory pressure.

Avoid using fielddata on text fields

For high-cardinality text fields, fielddata can use a large amount of JVM memory. To avoid this, {es} disables fielddata on text fields by default. If you’ve enabled fielddata and triggered the fielddata circuit breaker, consider disabling it and using a keyword field instead. See [fielddata].

Clear the fieldata cache

If you’ve triggered the fielddata circuit breaker and can’t disable fielddata, use the clear cache API to clear the fielddata cache. This may disrupt any in-flight searches that use fielddata.

POST _cache/clear?fielddata=true

High CPU usage

{es} uses thread pools to manage CPU resources for concurrent operations. High CPU usage typically means one or more thread pools are running low.

If a thread pool is depleted, {es} will reject requests related to the thread pool. For example, if the search thread pool is depleted, {es} will reject search requests until more threads are available.

Diagnose high CPU usage

Check CPU usage

Check hot threads

If a node has high CPU usage, use the nodes hot threads API to check for resource-intensive threads running on the node.

GET _nodes/my-node,my-other-node/hot_threads

This API returns a breakdown of any hot threads in plain text.

Reduce CPU usage

The following tips outline the most common causes of high CPU usage and their solutions.

Scale your cluster

Heavy indexing and search loads can deplete smaller thread pools. To better handle heavy workloads, add more nodes to your cluster or upgrade your existing nodes to increase capacity.

Spread out bulk requests

While more efficient than individual requests, large bulk indexing or multi-search requests still require CPU resources. If possible, submit smaller requests and allow more time between them.

Cancel long-running searches

Long-running searches can block threads in the search thread pool. To check for these searches, use the task management API.

GET _tasks?actions=*search&detailed

The response’s description contains the search request and its queries. running_time_in_nanos shows how long the search has been running.

{
  "nodes" : {
    "oTUltX4IQMOUUVeiohTt8A" : {
      "name" : "my-node",
      "transport_address" : "127.0.0.1:9300",
      "host" : "127.0.0.1",
      "ip" : "127.0.0.1:9300",
      "tasks" : {
        "oTUltX4IQMOUUVeiohTt8A:464" : {
          "node" : "oTUltX4IQMOUUVeiohTt8A",
          "id" : 464,
          "type" : "transport",
          "action" : "indices:data/read/search",
          "description" : "indices[my-index], search_type[QUERY_THEN_FETCH], source[{\"query\":...}]",
          "start_time_in_millis" : 4081771730000,
          "running_time_in_nanos" : 13991383,
          "cancellable" : true
        }
      }
    }
  }
}

To cancel a search and free up resources, use the API’s _cancel endpoint.

POST _tasks/oTUltX4IQMOUUVeiohTt8A:464/_cancel

For additional tips on how to track and avoid resource-intensive searches, see Avoid expensive searches.

High JVM memory pressure

High JVM memory usage can degrade cluster performance and trigger circuit breaker errors. To prevent this, we recommend taking steps to reduce memory pressure if a node’s JVM memory usage consistently exceeds 85%.

Diagnose high JVM memory pressure

Check JVM memory pressure

Check garbage collection logs

As memory usage increases, garbage collection becomes more frequent and takes longer. You can track the frequency and length of garbage collection events in elasticsearch.log. For example, the following event states {es} spent more than 50% (21 seconds) of the last 40 seconds performing garbage collection.

[timestamp_short_interval_from_last][INFO ][o.e.m.j.JvmGcMonitorService] [node_id] [gc][number] overhead, spent [21s] collecting in the last [40s]

Reduce JVM memory pressure

Reduce your shard count

Every shard uses memory. In most cases, a small set of large shards uses fewer resources than many small shards. For tips on reducing your shard count, see [size-your-shards].

Avoid expensive searches

Expensive searches can use large amounts of memory. To better track expensive searches on your cluster, enable slow logs.

Expensive searches may have a large size argument, use aggregations with a large number of buckets, or include expensive queries. To prevent expensive searches, consider the following setting changes:

PUT _settings
{
  "index.max_result_window": 5000
}

PUT _cluster/settings
{
  "persistent": {
    "search.max_buckets": 20000,
    "search.allow_expensive_queries": false
  }
}

Prevent mapping explosions

Defining too many fields or nesting fields too deeply can lead to mapping explosions that use large amounts of memory. To prevent mapping explosions, use the mapping limit settings to limit the number of field mappings.

Spread out bulk requests

While more efficient than individual requests, large bulk indexing or multi-search requests can still create high JVM memory pressure. If possible, submit smaller requests and allow more time between them.

Upgrade node memory

Heavy indexing and search loads can cause high JVM memory pressure. To better handle heavy workloads, upgrade your nodes to increase their memory capacity.

Red or yellow cluster status

A red or yellow cluster status indicates one or more shards are missing or unallocated. These unassigned shards increase your risk of data loss and can degrade cluster performance.

Diagnose your cluster status

Check your cluster status

GET _cluster/health?filter_path=status,*_shards

A healthy cluster has a green status and zero unassigned_shards. A yellow status means only replicas are unassigned. A red status means one or more primary shards are unassigned.

View unassigned shards

To view unassigned shards, use the cat shards API.

GET _cat/shards?v=true&h=index,shard,prirep,state,node,unassigned.reason&s=state

Unassigned shards have a state of UNASSIGNED. The prirep value is p for primary shards and r for replicas.

To understand why an unassigned shard is not being assigned and what action you must take to allow {es} to assign it, use the cluster allocation explanation API.

GET _cluster/allocation/explain?filter_path=index,node_allocation_decisions.node_name,node_allocation_decisions.deciders.*
{
  "index": "my-index",
  "shard": 0,
  "primary": false,
  "current_node": "my-node"
}

Fix a red or yellow cluster status

A shard can become unassigned for several reasons. The following tips outline the most common causes and their solutions.

Re-enable shard allocation

You typically disable allocation during a restart or other cluster maintenance. If you forgot to re-enable allocation afterward, {es} will be unable to assign shards. To re-enable allocation, reset the cluster.routing.allocation.enable cluster setting.

PUT _cluster/settings
{
  "persistent" : {
    "cluster.routing.allocation.enable" : null
  }
}

Recover lost nodes

Shards often become unassigned when a data node leaves the cluster. This can occur for several reasons, ranging from connectivity issues to hardware failure. After you resolve the issue and recover the node, it will rejoin the cluster. {es} will then automatically allocate any unassigned shards.

To avoid wasting resources on temporary issues, {es} delays allocation by one minute by default. If you’ve recovered a node and don’t want to wait for the delay period, you can call the cluster reroute API with no arguments to start the allocation process. The process runs asynchronously in the background.

POST _cluster/reroute

Fix allocation settings

Misconfigured allocation settings can result in an unassigned primary shard. These settings include:

To review your allocation settings, use the get index settings and cluster get settings APIs.

GET my-index/_settings?flat_settings=true&include_defaults=true

GET _cluster/settings?flat_settings=true&include_defaults=true

You can change the settings using the update index settings and cluster update settings APIs.

Allocate or reduce replicas

To protect against hardware failure, {es} will not assign a replica to the same node as its primary shard. If no other data nodes are available to host the replica, it remains unassigned. To fix this, you can:

  • Add a data node to the same tier to host the replica.

  • Change the index.number_of_replicas index setting to reduce the number of replicas for each primary shard. We recommend keeping at least one replica per primary.

PUT _settings
{
  "index.number_of_replicas": 1
}

Free up or increase disk space

{es} uses a low disk watermark to ensure data nodes have enough disk space for incoming shards. By default, {es} does not allocate shards to nodes using more than 85% of disk space.

To check the current disk space of your nodes, use the cat allocation API.

GET _cat/allocation?v=true&h=node,shards,disk.*

If your nodes are running low on disk space, you have a few options:

  • Upgrade your nodes to increase disk space.

  • Delete unneeded indices to free up space. If you use {ilm-init}, you can update your lifecycle policy to use searchable snapshots or add a delete phase. If you no longer need to search the data, you can use a snapshot to store it off-cluster.

  • If you no longer write to an index, use the force merge API or {ilm-init}'s force merge action to merge its segments into larger ones.

    POST my-index/_forcemerge

  • If an index is read-only, use the shrink index API or {ilm-init}'s shrink action to reduce its primary shard count.

    POST my-index/_shrink/my-shrunken-index

  • If your node has a large disk capacity, you can increase the low disk watermark or set it to an explicit byte value.

    PUT _cluster/settings
{
  "persistent": {
    "cluster.routing.allocation.disk.watermark.low": "30gb"
  }
}

Reduce JVM memory pressure

Shard allocation requires JVM heap memory. High JVM memory pressure can trigger circuit breakers that stop allocation and leave shards unassigned. See High JVM memory pressure.

Recover data for a lost primary shard

If a node containing a primary shard is lost, {es} can typically replace it using a replica on another node. If you can’t recover the node and replicas don’t exist or are irrecoverable, you’ll need to re-add the missing data from a snapshot or the original data source.

Warning
 Only use this option if node recovery is no longer possible. This process allocates an empty primary shard. If the node later rejoins the cluster, {es} will overwrite its primary shard with data from this newer empty shard, resulting in data loss.

Use the cluster reroute API to manually allocate the unassigned primary shard to another data node in the same tier. Set accept_data_loss to true.

POST _cluster/reroute
{
  "commands": [
    {
      "allocate_empty_primary": {
        "index": "my-index",
        "shard": 0,
        "node": "my-node",
        "accept_data_loss": "true"
      }
    }
  ]
}

If you backed up the missing index data to a snapshot, use the restore snapshot API to restore the individual index. Alternatively, you can index the missing data from the original data source.

Rejected requests

When {es} rejects a request, it stops the operation and returns an error with a 429 response code. Rejected requests are commonly caused by:

Check rejected tasks

To check the number of rejected tasks for each thread pool, use the cat thread pool API. A high ratio of rejected to completed tasks, particularly in the search and write thread pools, means {es} regularly rejects requests.

GET /_cat/thread_pool?v=true&h=id,name,active,rejected,completed

Prevent rejected requests

Fix high CPU and memory usage

If {es} regularly rejects requests and other tasks, your cluster likely has high CPU usage or high JVM memory pressure. For tips, see High CPU usage and High JVM memory pressure.

Prevent circuit breaker errors

If you regularly trigger circuit breaker errors, see Circuit breaker errors for tips on diagnosing and preventing them.

Task queue backlog

A backlogged task queue can prevent tasks from completing and put the cluster into an unhealthy state. Resource constraints, a large number of tasks being triggered at once, and long running tasks can all contribute to a backlogged task queue.

Diagnose a task queue backlog

Check the thread pool status

You can use the cat thread pool API to see the number of active threads in each thread pool and how many tasks are queued, how many have been rejected, and how many have completed.

GET /_cat/thread_pool?v&s=t,n&h=type,name,node_name,active,queue,rejected,completed

Inspect the hot threads on each node

If a particular thread pool queue is backed up, you can periodically poll the Nodes hot threads API to determine if the thread has sufficient resources to progress and gauge how quickly it is progressing.

GET /_nodes/hot_threads

Look for long running tasks

Long-running tasks can also cause a backlog. You can use the task management API to get information about the tasks that are running. Check the running_time_in_nanos to identify tasks that are taking an excessive amount of time to complete.

GET /_tasks?filter_path=nodes.*.tasks

Resolve a task queue backlog

Increase available resources

If tasks are progressing slowly and the queue is backing up, you might need to take steps to Reduce CPU usage.

In some cases, increasing the thread pool size might help. For example, the force_merge thread pool defaults to a single thread. Increasing the size to 2 might help reduce a backlog of force merge requests.

Cancel stuck tasks

If you find the active task’s hot thread isn’t progressing and there’s a backlog, consider canceling the task.