Migrate from InfluxDB to VictoriaMetrics

InfluxDB is a well-known time series database built for IoT monitoring, Application Performance Monitoring (APM) and analytics. It has its query language, unique data model, and rich tooling for collecting and processing metrics.

Nowadays, the volume of time series data grows constantly, as well as requirements for durable time series storage. And sometimes old known solutions just can’t keep up with the new expectations.

VictoriaMetrics is a high-performance opensource time series database specifically designed to deal with huge volumes of monitoring data while remaining cost-efficient at the same time. Many companies are choosing to migrate from InfluxDB to VictoriaMetrics specifically for performance and scalability reasons. Along them see case studies provided by ARNES and Brandwatch.

This guide will cover the differences between two solutions, most commonly asked questions, and approaches for migrating from InfluxDB to VictoriaMetrics.

Data model differences#

While readers are likely familiar with InfluxDB key concepts, the data model of VictoriaMetrics is something new to explore. Let’s start with similarities and differences:

• both solutions are schemaless, which means there is no need to define metrics or their tags in advance;
• multidimensional data support is implemented via tags in InfluxDB and via labels in VictoriaMetrics. However, labels in VictoriaMetrics are always strings, while InfluxDB supports multiple data types;
• timestamps are stored with nanosecond resolution in InfluxDB, while in VictoriaMetrics it is milliseconds;
• in VictoriaMetrics metric value is always float64, while InfluxDB supports multiple data types.
• there are no measurements or fields in VictoriaMetrics, metric name contains it all. If measurement contains more than 1 field, then for VictoriaMetrics it will be multiple metrics;
• there are no databases, buckets or organizations. All data in VictoriaMetrics is stored in a global namespace or within a tenant. See more about multi-tenancy here.

Let’s consider the following sample data borrowed from InfluxDB docs as an example:

In VictoriaMetrics data model this sample will have the following form:

Actually, metric name for VictoriaMetrics is also a label with static name __name__, and example above can be converted to {__name__="census_bees", location="klamath", scientist="anderson"}. All labels are indexed by VictoriaMetrics, so lookups by names or labels have the same query speed.

Write data#

VictoriaMetrics supports InfluxDB line protocol for data ingestion. For example, to write a measurement to VictoriaMetrics we need to send an HTTP POST request with payload in a line protocol format:

curl -d 'census,location=klamath,scientist=anderson bees=23 1566079200000' -X POST 'http://<victoriametric-addr>:8428/write'

hint: timestamp in the example might be out of configured retention for VictoriaMetrics. Consider increasing the retention period or changing the timestamp, if that is the case.

Please note, an arbitrary number of lines delimited by \n (aka newline char) can be sent in a single request.

To get the written data back let’s export all series matching the location="klamath" filter:

curl -G 'http://<victoriametric-addr>:8428/api/v1/export' -d 'match={location="klamath"}'   

The expected response is the following:

{
  "metric": {
    "__name__": "census_bees",
    "location": "klamath",
    "scientist": "anderson"
  },
  "values": [
    23
  ],
  "timestamps": [
    1566079200000
  ]
}

Please note, VictoriaMetrics performed additional data mapping to the data ingested via InfluxDB line protocol.

Support of InfluxDB line protocol also means VictoriaMetrics is compatible with Telegraf. To configure Telegraf, simply add http://<victoriametric-addr>:8428 URL to Telegraf configs:

[[outputs.influxdb]]
  urls = ["http://<victoriametrics-addr>:8428"]

In addition to InfluxDB line protocol, VictoriaMetrics supports many other ways for metrics collection.

Query data#

VictoriaMetrics does not have a command-line interface (CLI). Instead, it provides an HTTP API for serving read queries. This API is used in various integrations such as Grafana. The same API is also used by VMUI - a graphical User Interface for querying and visualizing metrics:

See more about how to query data in VictoriaMetrics.

Basic concepts#

Let’s take a closer look at querying specific with the following data sample:

foo
,instance=localhost bar=1.00 1652169600000000000
foo,instance=localhost bar=2.00 1652169660000000000
foo,instance=localhost bar=3.00 1652169720000000000
foo,instance=localhost bar=5.00 1652169840000000000
foo,instance=localhost bar=5.50 1652169960000000000
foo,instance=localhost bar=5.50 1652170020000000000
foo,instance=localhost bar=4.00 1652170080000000000
foo,instance=localhost bar=3.50 1652170260000000000
foo,instance=localhost bar=3.25 1652170320000000000
foo,instance=localhost bar=3.00 1652170380000000000
foo,instance=localhost bar=2.00 1652170440000000000
foo,instance=localhost bar=1.00 1652170500000000000
foo,instance=localhost bar=4.00 1652170560000000000

The data sample consists data points for a measurement foo and a field bar with additional tag instance=localhost. If we would like plot this data as a time series in Grafana it might have the following look:

The query used for this panel is written in InfluxQL:

SELECT last ("bar")
FROM "foo"
WHERE ("instance" = 'localhost')
  AND $timeFilter
GROUP BY time (1m)

Having this, let’s import the same data sample in VictoriaMetrics and plot it in Grafana as well. To understand how the InfluxQL query might be translated to MetricsQL let’s break it into components first:

• SELECT last("bar") FROM "foo" - all requests to instant or range VictoriaMetrics APIs are reads, so no need to specify the SELECT statement. There are no measurements or fields in VictoriaMetrics, so the whole expression can be replaced with foo_bar in MetricsQL;
• WHERE ("instance" = 'localhost')- filtering by labels in MetricsQL requires specifying the filter in curly braces next to the metric name. So in MetricsQL filter expression will be translated to {instance="localhost"};
• WHERE $timeFilter - filtering by time is done via request params sent along with query, so in MetricsQL no need to specify this filter;
• GROUP BY time(1m) - grouping by time is done by default in range API according to specified step param. This param is also a part of params sent along with request. See how to perform additional aggregations and grouping via MetricsQL .

In result, executing the foo_bar{instance="localhost"} MetricsQL expression with step=1m for the same set of data in Grafana will have the following form:

Visualizations from both databases are a bit different - VictoriaMetrics shows some extra points filling the gaps in the graph. This behavior is described in more detail here. In InfluxDB, we can achieve a similar behavior by adding fill(previous) to the query.

VictoriaMetrics fills the gaps on the graph assuming time series are always continuous and not discrete. To limit the interval on which VictoriaMetrics will try to fill the gaps, set -search.setLookbackToStep command-line flag. This limits the gap filling to a single step interval passed to /api/v1/query_range. This behavior is close to InfluxDB data model.

Advanced usage#

The good thing is that knowing the basics and some aggregation functions is often enough for using MetricsQL or PromQL. Let’s consider one of the most popular Grafana dashboards Node Exporter Full. It has almost 15 million downloads and about 230 PromQL queries in it! But a closer look at those queries shows the following:

• ~120 queries are just selecting a metric with label filters, e.g. node_textfile_scrape_error{instance="$node",job="$job"};
• ~80 queries are using rate function for selected metric, e.g. rate(node_netstat_Tcp_InSegs{instance=\"$node\",job=\"$job\"})
• and the rest are aggregation functions like sum or count.

To get a better understanding of how MetricsQL works, see the following resources:

• MetricsQL concepts;
• MetricsQL functions;
• PromQL tutorial for beginners.

How to migrate current data from InfluxDB to VictoriaMetrics#

Migrating data from other TSDBs to VictoriaMetrics is as simple as importing data via any of supported formats.

But migration from InfluxDB might get easier when using vmctl - VictoriaMetrics command-line tool. See more about migrating from InfluxDB v1.x versions. Migrating data from InfluxDB v2.x is not supported yet. But there is useful 3rd party solution for this.

Please note, that data migration is a backfilling process. So, please consider backfilling tips.

Frequently asked questions#

• How does VictoriaMetrics compare to InfluxDB?
  • Answer
• Why don’t VictoriaMetrics support Remote Read API, so I don’t need to learn MetricsQL?
  • Answer
• The PromQL and MetricsQL are often mentioned together - why is that?
  • MetricsQL - query language inspired by PromQL. MetricsQL is backward-compatible with PromQL, so Grafana dashboards backed by Prometheus datasource should work the same after switching from Prometheus to VictoriaMetrics. Both languages mostly share the same concepts with slight differences.
• Query returns more data points than expected - why?
  • VictoriaMetrics may return non-existing data points if step param is lower than the actual data resolution. See more about this here.
• How do I get the real last data point, not ephemeral?
  • last_over_time function can be used for limiting the lookbehind window for calculated data. For example, last_over_time(metric[10s]) would return calculated samples only if the real samples are located closer than 10 seconds to the calculated timestamps according to start, end and step query args passed to range query.
• How do I get raw data points with MetricsQL?
  • For getting raw data points specify the interval at which you want them in square brackets and send as instant query. For example, GET api/v1/query?query=my_metric[5m]&time=<time> will return raw samples for my_metric in interval from <time> to <time>-5m.
• Can you have multiple aggregators in a MetricsQL query, e.g. SELECT MAX(field), MIN(field) ...?
  • Yes, try the following query ( alias(max(field), "max"), alias(min(field), "min") ).
• How to translate Influx percentile function to MetricsQL?
  • Answer
• How to translate Influx stddev function to MetricsQL?
  • Answer