Skip to content

report-query-performance-issues.md

Latest commit

 

History

History
443 lines (344 loc) · 15.2 KB

report-query-performance-issues.md

File metadata and controls

443 lines (344 loc) · 15.2 KB
Raw
title Report query performance issues
description Follow guidelines to help InfluxData engineers troubleshoot and resolve query performance issues.
menu
influxdb3_clustered
name parent
Report query performance issues
Troubleshoot and optimize queries
weight 402
related
/influxdb3/clustered/admin/query-system-data/
/influxdb3/clustered/query-data/troubleshoot-and-optimize/analyze-query-plan/

Use these guidelines to work with InfluxData engineers to troubleshoot and resolve query performance issues.

Note

Optimize your query

Before reporting a query performance problem, see the troubleshooting and optimization guide to learn how to optimize your query and reduce compute and memory requirements.

  1. Send InfluxData output artifacts
  2. Document your test process
  3. Document your environment
  4. Document your data schema
  5. Establish query performance degradation conditions
  6. Reduce query noise
  7. Establish baseline single-query performance
  8. Run queries at multiple load scales
  9. Gather debug information
    1. Kubernetes-specific information
    2. Clustered-specific information
    3. Query analysis
      1. EXPLAIN
      2. EXPLAIN VERBOSE
      3. EXPLAIN ANALYZE
  10. Gather system information

Note

Please note that this document may change from one support engagement to the next as our process and tooling improves.

Send InfluxData output artifacts

As you follow these guidelines, package all produced output artifacts in the following form:

Outputs:

  • test-artifact-name.tar.gz

Send InfluxData engineers all produced artifacts for analysis.

Document your test process

Currently, {{% product-name %}} doesn't provide a standardized performance test suite that you can run in your cluster. Please document your test process so that InfluxData engineers can replicate it--include the following:

  • The steps you take when performance testing.
  • Timestamps of your test runs, to correlate tests with logs.

Document your environment

Provide as much detail about your environment as your organization allows, including the following:

  • Your kubernetes cluster
  • The cloud provider where it runs or indicate that it's "on-prem"
  • The hardware it runs on
  • The type and size of disk in use--for example: hard disk, SSD, NVMe, etc.
  • CPU and memory resources set on each type of InfluxDB pod
  • The number of pods in each InfluxDB StatefulSet and Deployment
  • The type of object store used and how it is hosted
  • How the Catalog store (PostgreSQL-compatible database) is hosted
  • Indicate if either the Object store or the Catalog store is shared by more than one InfluxDB Clustered product
    • If so, describe the network-level topology of your setup

Note

If possible, provide a synthetic dataset

If you can reproduce the performance issue with a synthetic dataset, and your process and environment are well-documented, InfluxData engineers may be able to reproduce the issue, shorten the feedback cycle, and resolve the issue sooner.

Document your data schema

Document your the data schema to help InfluxData engineers better understand the conditions that reproduce your issue.

Establish query performance degradation conditions

The most effective way to investigate query performance is to have a good understanding of the conditions in which you don't see the expected performance. Consider the following:

  • Does this always happen, or only sometimes?
    • If only sometimes, is it at a consistent time of day or over a consistent period?
  • Will a single query execution reproduce the issue, or does it only appear with multiple queries are running at the same time?
  • How are you executing the queries? For example:
    • influxctl
    • Client libraries
    • Other environments or tools

Reduce query noise

Test in an environment without periodic or intermittent queries to measure baseline system performance without additional query noise.

When running multiple tests with different queries, allow the system to recover between tests. Wait at least one minute after receiving a query result before executing the next query.

Establish baseline single-query performance

Perform some tests with single queries in isolation to measure baseline performance. This approach may not always reproduce your issue but can provide useful data for analysis by InfluxData engineers.

Run queries at multiple load scales

If the issue isn't replicated after reducing query noise and establishing baseline single-query performance, systematically increase query concurrency to reproduce the problem and identify the scale at which it occurs--for example, run the following test plan.

Note

You might need to scale the example plan up or down, as necessary, to reproduce the problem.

  1. Turn off intermittent or periodic InfluxDB queries and allow the cluster to recover.
  2. Run Query A and allow the cluster to recover for 1 minute.
  3. Run 5 concurrent instances of Query A and allow the cluster to recover for 1 minute.
  4. Run 10 concurrent instances of Query A and allow the cluster to recover for 1 minute.
  5. Run 20 concurrent instances of Query A and allow the cluster to recover for 1 minute.
  6. Run 40 concurrent instances of Query A and allow the cluster to recover for 1 minute.
  7. Provide InfluxData the debug information associated with each test run.

Your test findings and associated debug information from your Kubernetes environment can help recommend configuration changes to improve query performance as your usage scales.

Capture dashboard screenshots

For query performance issues, always capture screenshots of the Querier Dashboard as a first step.

If you have set up alerts and dashboards for monitoring your cluster, capture screenshots of dashboard events for Queriers, Compactors, and Ingesters.

On the Querier dashboard, capture screenshots showing:

  • CPU utilization: Is it running high (close to the limits you set)?
  • Object Store Traffic/Latency: Often a major contributor to performance issues
  • Cache Requests bytes: Shows cache misses as separate series
  • Query concurrency and rate metrics:
    • grpc Requests
    • Query Rate
    • Query Concurrency (note the 10-minute maximum limitation)
  • Parquet files per query: Number of files accessed per query
  • Request Duration...DoGet: Query execution timing

Gather debug information

Shortly after testing a problematic query against your InfluxDB cluster, collect the following debug information.

Kubernetes-specific information

Outputs:

  • ${DATETIME}-cluster-info.tar.gz
DATETIME="$(date -Iminutes)"
kubectl cluster-info dump --namespace influxdb --output-directory "${DATETIME}-cluster-info/"
tar -czf "${DATETIME}-cluster-info.tar.gz" "${DATETIME}-cluster-info/"

Clustered-specific information

Outputs:

  • app-instance.yml: Provide a copy of your AppInstance manifest.

Query analysis

Use EXPLAIN commands to output query plan information for a long-running query.

Outputs (InfluxQL):

  • explain.csv
  • explain-verbose.csv
  • explain-analyze.csv

Outputs (SQL):

  • explain.txt
  • explain-verbose.txt
  • explain-analyze.txt

In the examples below, replace the following:

  • {{% code-placeholder-key %}}DATABASE_NAME{{% /code-placeholder-key %}}: The name of the database to query
  • {{% code-placeholder-key %}}DATABASE_TOKEN{{% /code-placeholder-key %}}: A database token with read permissions on the queried database
  • {{% code-placeholder-key %}}YOUR_QUERY{{% /code-placeholder-key %}}: Your long-running query (formatted as a single line with escaped double quotes (\"))
EXPLAIN

{{% code-placeholders "DATABASE_(NAME|TOKEN)|YOUR_QUERY" %}}

{{< code-tabs-wrapper >}} {{% code-tabs %}} SQL InfluxQL {{% /code-tabs %}} {{% code-tab-content %}}

influxctl \
  --config config.toml \
    query \
  --database DATABASE_NAME \
  --format table \
  --token DATABASE_TOKEN \
  "EXPLAIN YOUR_QUERY;" > explain.txt

{{% /code-tab-content %}} {{% code-tab-content %}}

curl --get "https://{{< influxdb/host >}}/query" \
  --output "./explain.csv" \
  --header "Authorization: Bearer DATABASE_TOKEN" \
  --header "Accept: application/csv" \
  --data-urlencode "db=DATABASE_NAME" \
  --data-urlencode "q=EXPLAIN YOUR_QUERY"

{{% /code-tab-content %}} {{< /code-tabs-wrapper >}}

{{% /code-placeholders %}}

EXPLAIN VERBOSE

{{% code-placeholders "DATABASE_(NAME|TOKEN)|YOUR_QUERY" %}}

{{< code-tabs-wrapper >}} {{% code-tabs %}} SQL InfluxQL {{% /code-tabs %}} {{% code-tab-content %}}

influxctl \
  --config config.toml \
    query \
  --database DATABASE_NAME \
  --format table \
  --token DATABASE_TOKEN \
  "EXPLAIN VERBOSE YOUR_QUERY;" > explain-verbose.txt

{{% /code-tab-content %}} {{% code-tab-content %}}

curl --get "https://{{< influxdb/host >}}/query" \
  --output "./explain-verbose.csv" \
  --header "Authorization: Bearer DATABASE_TOKEN" \
  --header "Accept: application/csv" \
  --data-urlencode "db=DATABASE_NAME" \
  --data-urlencode "q=EXPLAIN VERBOSE YOUR_QUERY"

{{% /code-tab-content %}} {{< /code-tabs-wrapper >}}

{{% /code-placeholders %}}

EXPLAIN ANALYZE

{{% code-placeholders "DATABASE_(NAME|TOKEN)|YOUR_QUERY" %}}

{{< code-tabs-wrapper >}} {{% code-tabs %}} SQL InfluxQL {{% /code-tabs %}} {{% code-tab-content %}}

influxctl \
  --config config.toml \
    query \
  --database DATABASE_NAME \
  --format table \
  --token DATABASE_TOKEN \
  "EXPLAIN ANALYZE YOUR_QUERY;" > explain-analyze.txt

{{% /code-tab-content %}} {{% code-tab-content %}}

curl --get "https://{{< influxdb/host >}}/query" \
  --output "./explain-analyze.csv" \
  --header "Authorization: Bearer DATABASE_TOKEN" \
  --header "Accept: application/csv" \
  --data-urlencode "db=DATABASE_NAME" \
  --data-urlencode "q=EXPLAIN ANALYZE YOUR_QUERY"

{{% /code-tab-content %}} {{< /code-tabs-wrapper >}}

{{% /code-placeholders %}}

Include `EXPLAIN ANALYZE` output.

When using the output for troubleshooting performance, focus on the sections with the highest `elapsed_compute` times, as these indicate performance bottlenecks.
For example, here is extracted timing data from an ANALYZE output showing the most time-consuming operations:

```text
DeduplicateExec
  └→ elapsed_compute=3.514663491s                                           3514.66ms
SortPreservingMergeExec
  └→ elapsed_compute=12.440516244s                                         12440.52ms
SortExec
  └→ elapsed_compute=993.952663ms                                            993.95ms
AggregateExec
  └→ elapsed_compute=406.163116ms                                            406.16ms
ParquetExec
  └→ time_elapsed_scanning_total=1044.149737489s                         1044149.74ms
  └→ time_elapsed_opening=3.001925899s                                      3001.93ms
  └→ time_elapsed_processing=2.255025048s                                   2255.03ms

Gather system information

Warning

May impact cluster performance

Querying InfluxDB 3 system tables may impact write and query performance of your {{< product-name omit=" Clustered" >}} cluster. Use filters to optimize queries to reduce impact to your cluster.

System tables are subject to change

System tables are not part of InfluxDB's stable API and may change with new releases. The provided schema information and query examples are valid as of September 20, 2024. If you detect a schema change or a non-functioning query example, please submit an issue.

If queries are slow for a specific table, run the following system queries to collect information for troubleshooting:

To optimize system queries, use table_name, partition_key, and partition_id filters. In your queries, replace the following:

  • {{% code-placeholder-key %}}TABLE_NAME{{% /code-placeholder-key %}}: the table to retrieve partitions for
  • {{% code-placeholder-key %}}PARTITION_ID{{% /code-placeholder-key %}}: a partition ID (int64)
  • {{% code-placeholder-key %}}PARTITION_KEY{{% /code-placeholder-key %}}: a partition key derived from the table's partition template. The default format is %Y-%m-%d (for example, 2024-01-01).

Collect table information

{{% code-placeholders "TABLE_NAME" %}}

SELECT *
FROM system.tables
WHERE table_name = 'TABLE_NAME';

{{% /code-placeholders%}}

Collect compaction information for the table

Query the system.compactor table to collect compaction information--for example, run one of the following queries:

{{% code-placeholders "TABLE_NAME|PARTITION_KEY" %}}

SELECT * 
FROM system.compactor 
WHERE
  table_name = 'TABLE_NAME' 
    AND partition_key = 'PARTITION_KEY';

{{% /code-placeholders %}}

{{% code-placeholders "TABLE_NAME|PARTITION_ID" %}}

SELECT * 
FROM system.compactor 
WHERE
  table_name = 'TABLE_NAME' 
    AND partition_id = 'PARTITION_ID';

{{% /code-placeholders %}}

Collect partition information for multiple tables

If the same queries are slow on more than 1 table, also run the following query to collect the size and number of partitions for all tables:

{{% code-placeholders "TABLE_NAME" %}}

SELECT table_name,
  COUNT(*) as partition_count,
  MAX(last_new_file_created_at) as last_new_file_created_at,
  SUM(total_size_mb) as total_size_mb
FROM system.partitions
WHERE table_name IN ('foo', 'bar', 'baz')
GROUP BY table_name;

{{% /code-placeholders%}}