elasticsearch/docs/reference/troubleshooting/common-issues/hotspotting.asciidoc
Arianna Laudazzi 70e5a67904
[AutoOps] Reference AutoOps solution on troubleshooting pages (#119630)
* Reference AutoOps on troubleshooting pages

* Integrate reviewer's feedback
2025-01-09 16:24:20 +01:00

286 lines
12 KiB
Text

[[hotspotting]]
=== Hot spotting
++++
<titleabbrev>Hot spotting</titleabbrev>
++++
:keywords: hot-spotting, hotspot, hot-spot, hot spot, hotspots, hotspotting
Computer link:{wikipedia}/Hot_spot_(computer_programming)[hot spotting]
may occur in {es} when resource utilizations are unevenly distributed across
<<modules-node,nodes>>. Temporary spikes are not usually considered problematic, but
ongoing significantly unique utilization may lead to cluster bottlenecks
and should be reviewed.
****
If you're using Elastic Cloud Hosted, then you can use AutoOps to monitor your cluster. AutoOps significantly simplifies cluster management with performance recommendations, resource utilization visibility, real-time issue detection and resolution paths. For more information, refer to https://www.elastic.co/guide/en/cloud/current/ec-autoops.html[Monitor with AutoOps].
****
See link:https://www.youtube.com/watch?v=Q5ODJ5nIKAM[this video] for a walkthrough of troubleshooting a hot spotting issue.
[discrete]
[[detect]]
==== Detect hot spotting
Hot spotting most commonly surfaces as significantly elevated
resource utilization (of `disk.percent`, `heap.percent`, or `cpu`) among a
subset of nodes as reported via <<cat-nodes,cat nodes>>. Individual spikes aren't
necessarily problematic, but if utilization repeatedly spikes or consistently remains
high over time (for example longer than 30 seconds), the resource may be experiencing problematic
hot spotting.
For example, let's show case two separate plausible issues using cat nodes:
[source,console]
----
GET _cat/nodes?v&s=master,name&h=name,master,node.role,heap.percent,disk.used_percent,cpu
----
Pretend this same output pulled twice across five minutes:
[source,console-result]
----
name master node.role heap.percent disk.used_percent cpu
node_1 * hirstm 24 20 95
node_2 - hirstm 23 18 18
node_3 - hirstmv 25 90 10
----
// TEST[skip:illustrative response only]
Here we see two significantly unique utilizations: where the master node is at
`cpu: 95` and a hot node is at `disk.used_percent: 90%`. This would indicate
hot spotting was occurring on these two nodes, and not necessarily from the same
root cause.
[discrete]
[[causes]]
==== Causes
Historically, clusters experience hot spotting mainly as an effect of hardware,
shard distributions, and/or task load. We'll review these sequentially in order
of their potentially impacting scope.
[discrete]
[[causes-hardware]]
==== Hardware
Here are some common improper hardware setups which may contribute to hot
spotting:
* Resources are allocated non-uniformly. For example, if one hot node is
given half the CPU of its peers. {es} expects all nodes on a
<<data-tiers,data tier>> to share the same hardware profiles or
specifications.
* Resources are consumed by another service on the host, including other
{es} nodes. Refer to our <<dedicated-host,dedicated host>> recommendation.
* Resources experience different network or disk throughputs. For example, if one
node's I/O is lower than its peers. Refer to
<<tune-for-indexing-speed,Use faster hardware>> for more information.
* A JVM that has been configured with a heap larger than 31GB. Refer to <<set-jvm-heap-size>>
for more information.
* Problematic resources uniquely report <<setup-configuration-memory,memory swapping>>.
[discrete]
[[causes-shards]]
==== Shard distributions
{es} indices are divided into one or more link:{wikipedia}/Shard_(database_architecture)[shards]
which can sometimes be poorly distributed. {es} accounts for this by <<modules-cluster,balancing shard counts>>
across data nodes. As link:{blog-ref}whats-new-elasticsearch-kibana-cloud-8-6-0[introduced in version 8.6],
{es} by default also enables <<modules-cluster,desired balancing>> to account for ingest load.
A node may still experience hot spotting either due to write-heavy indices or by the
overall shards it's hosting.
[discrete]
[[causes-shards-nodes]]
===== Node level
You can check for shard balancing via <<cat-allocation,cat allocation>>, though as of version
8.6, <<modules-cluster,desired balancing>> may no longer fully expect to
balance shards. Kindly note, both methods may temporarily show problematic imbalance during
<<cluster-fault-detection,cluster stability issues>>.
For example, let's showcase two separate plausible issues using cat allocation:
[source,console]
----
GET _cat/allocation?v&s=node&h=node,shards,disk.percent,disk.indices,disk.used
----
Which could return:
[source,console-result]
----
node shards disk.percent disk.indices disk.used
node_1 446 19 154.8gb 173.1gb
node_2 31 52 44.6gb 372.7gb
node_3 445 43 271.5gb 289.4gb
----
// TEST[skip:illustrative response only]
Here we see two significantly unique situations. `node_2` has recently
restarted, so it has a much lower number of shards than all other nodes. This
also relates to `disk.indices` being much smaller than `disk.used` while shards
are recovering as seen via <<cat-recovery,cat recovery>>. While `node_2`'s shard
count is low, it may become a write hot spot due to ongoing <<ilm-rollover,ILM
rollovers>>. This is a common root cause of write hot spots covered in the next
section.
The second situation is that `node_3` has a higher `disk.percent` than `node_1`,
even though they hold roughly the same number of shards. This occurs when either
shards are not evenly sized (refer to <<shard-size-recommendation>>) or when
there are a lot of empty indices.
Cluster rebalancing based on desired balance does much of the heavy lifting
of keeping nodes from hot spotting. It can be limited by either nodes hitting
<<disk-based-shard-allocation,watermarks>> (refer to <<fix-watermark-errors,fixing disk watermark errors>>) or by a
write-heavy index's total shards being much lower than the written-to nodes.
You can confirm hot spotted nodes via <<cluster-nodes-stats,the nodes stats API>>,
potentially polling twice over time to only checking for the stats differences
between them rather than polling once giving you stats for the node's
full <<cluster-nodes-usage,node uptime>>. For example, to check all nodes
indexing stats:
[source,console]
----
GET _nodes/stats?human&filter_path=nodes.*.name,nodes.*.indices.indexing
----
[discrete]
[[causes-shards-index]]
===== Index level
Hot spotted nodes frequently surface via <<cat-thread-pool,cat thread pool>>'s
`write` and `search` queue backups. For example:
[source,console]
----
GET _cat/thread_pool/write,search?v=true&s=n,nn&h=n,nn,q,a,r,c
----
Which could return:
[source,console-result]
----
n nn q a r c
search node_1 3 1 0 1287
search node_2 0 2 0 1159
search node_3 0 1 0 1302
write node_1 100 3 0 4259
write node_2 0 4 0 980
write node_3 1 5 0 8714
----
// TEST[skip:illustrative response only]
Here you can see two significantly unique situations. Firstly, `node_1` has a
severely backed up write queue compared to other nodes. Secondly, `node_3` shows
historically completed writes that are double any other node. These are both
probably due to either poorly distributed write-heavy indices, or to multiple
write-heavy indices allocated to the same node. Since primary and replica writes
are majorly the same amount of cluster work, we usually recommend setting
<<total-shards-per-node,`index.routing.allocation.total_shards_per_node`>> to
force index spreading after lining up index shard counts to total nodes.
We normally recommend heavy-write indices have sufficient primary
`number_of_shards` and replica `number_of_replicas` to evenly spread across
indexing nodes. Alternatively, you can <<cluster-reroute,reroute>> shards to
more quiet nodes to alleviate the nodes with write hot spotting.
If it's non-obvious what indices are problematic, you can introspect further via
<<indices-stats,the index stats API>> by running:
[source,console]
----
GET _stats?level=shards&human&expand_wildcards=all&filter_path=indices.*.total.indexing.index_total
----
For more advanced analysis, you can poll for shard-level stats,
which lets you compare joint index-level and node-level stats. This analysis
wouldn't account for node restarts and/or shards rerouting, but serves as
overview:
[source,console]
----
GET _stats/indexing,search?level=shards&human&expand_wildcards=all
----
You can for example use the link:https://stedolan.github.io/jq[third-party JQ tool],
to process the output saved as `indices_stats.json`:
[source,sh]
----
cat indices_stats.json | jq -rc ['.indices|to_entries[]|.key as $i|.value.shards|to_entries[]|.key as $s|.value[]|{node:.routing.node[:4], index:$i, shard:$s, primary:.routing.primary, size:.store.size, total_indexing:.indexing.index_total, time_indexing:.indexing.index_time_in_millis, total_query:.search.query_total, time_query:.search.query_time_in_millis } | .+{ avg_indexing: (if .total_indexing>0 then (.time_indexing/.total_indexing|round) else 0 end), avg_search: (if .total_search>0 then (.time_search/.total_search|round) else 0 end) }'] > shard_stats.json
# show top written-to shard simplified stats which contain their index and node references
cat shard_stats.json | jq -rc 'sort_by(-.avg_indexing)[]' | head
----
[discrete]
[[causes-tasks]]
==== Task loads
Shard distribution problems will most-likely surface as task load as seen
above in the <<cat-thread-pool,cat thread pool>> example. It is also
possible for tasks to hot spot a node either due to
individual qualitative expensiveness or overall quantitative traffic loads.
For example, if <<cat-thread-pool,cat thread pool>> reported a high
queue on the `warmer` <<modules-threadpool,thread pool>>, you would
look-up the effected node's <<cluster-nodes-hot-threads,hot threads>>.
Let's say it reported `warmer` threads at `100% cpu` related to
`GlobalOrdinalsBuilder`. This would let you know to inspect
<<eager-global-ordinals,field data's global ordinals>>.
Alternatively, let's say <<cat-nodes,cat nodes>> shows a hot spotted master node
and <<cat-thread-pool,cat thread pool>> shows general queuing across nodes.
This would suggest the master node is overwhelmed. To resolve
this, first ensure <<high-availability-cluster-small-clusters,hardware high availability>>
setup and then look to ephemeral causes. In this example,
<<cluster-nodes-hot-threads,the nodes hot threads API>> reports multiple threads in
`other` which indicates they're waiting on or blocked by either garbage collection
or I/O.
For either of these example situations, a good way to confirm the problematic tasks
is to look at longest running non-continuous (designated `[c]`) tasks via
<<cat-tasks,cat task management>>. This can be supplemented checking longest
running cluster sync tasks via <<cat-pending-tasks,cat pending tasks>>. Using
a third example,
[source,console]
----
GET _cat/tasks?v&s=time:desc&h=type,action,running_time,node,cancellable
----
This could return:
[source,console-result]
----
type action running_time node cancellable
direct indices:data/read/eql 10m node_1 true
...
----
// TEST[skip:illustrative response only]
This surfaces a problematic <<eql-search-api,EQL query>>. We can gain
further insight on it via <<tasks,the task management API>>,
[source,console]
----
GET _tasks?human&detailed
----
Its response contains a `description` that reports this query:
[source,eql]
----
indices[winlogbeat-*,logs-window*], sequence by winlog.computer_name with maxspan=1m\n\n[authentication where host.os.type == "windows" and event.action:"logged-in" and\n event.outcome == "success" and process.name == "svchost.exe" ] by winlog.event_data.TargetLogonId
----
This lets you know which indices to check (`winlogbeat-*,logs-window*`), as well
as the <<eql-search-api,EQL search>> request body. Most likely this is
link:{security-guide}/es-overview.html[SIEM related].
You can combine this with <<enable-audit-logging,audit logging>> as needed to
trace the request source.