in [source code](https://github.com/VictoriaMetrics/VictoriaMetrics). Just download VictoriaMetrics and see [how to start it](#how-to-start-victoriametrics).
* VictoriaMetrics can be used as long-term storage for Prometheus or for [vmagent](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmagent/README.md).
* Supports [Prometheus querying API](https://prometheus.io/docs/prometheus/latest/querying/api/), so it can be used as Prometheus drop-in replacement in Grafana.
* Supports global query view. Multiple Prometheus instances may write data into VictoriaMetrics. Later this data may be used in a single query.
* High performance and good scalability for both [inserts](https://medium.com/@valyala/high-cardinality-tsdb-benchmarks-victoriametrics-vs-timescaledb-vs-influxdb-13e6ee64dd6b)
and [selects](https://medium.com/@valyala/when-size-matters-benchmarking-victoriametrics-vs-timescale-and-influxdb-6035811952d4).
[Outperforms InfluxDB and TimescaleDB by up to 20x](https://medium.com/@valyala/measuring-vertical-scalability-for-time-series-databases-in-google-cloud-92550d78d8ae).
* [Uses 10x less RAM than InfluxDB](https://medium.com/@valyala/insert-benchmarks-with-inch-influxdb-vs-victoriametrics-e31a41ae2893) when working with millions of unique time series (aka high cardinality).
* Optimized for time series with high churn rate. Think about [prometheus-operator](https://github.com/coreos/prometheus-operator) metrics from frequent deployments in Kubernetes.
* High data compression, so [up to 70x more data points](https://medium.com/@valyala/when-size-matters-benchmarking-victoriametrics-vs-timescale-and-influxdb-6035811952d4)
may be crammed into limited storage comparing to TimescaleDB.
* Optimized for storage with high-latency IO and low IOPS (HDD and network storage in AWS, Google Cloud, Microsoft Azure, etc). See [graphs from these benchmarks](https://medium.com/@valyala/high-cardinality-tsdb-benchmarks-victoriametrics-vs-timescaledb-vs-influxdb-13e6ee64dd6b).
* A single-node VictoriaMetrics may substitute moderately sized clusters built with competing solutions such as Thanos, M3DB, Cortex, InfluxDB or TimescaleDB.
See [vertical scalability benchmarks](https://medium.com/@valyala/measuring-vertical-scalability-for-time-series-databases-in-google-cloud-92550d78d8ae),
[comparing Thanos to VictoriaMetrics cluster](https://medium.com/@valyala/comparing-thanos-to-victoriametrics-cluster-b193bea1683)
and [Remote Write Storage Wars](https://promcon.io/2019-munich/talks/remote-write-storage-wars/) talk
from [PromCon 2019](https://promcon.io/2019-munich/talks/remote-write-storage-wars/).
* VictoriaMetrics consists of a single [small executable](https://medium.com/@valyala/stripping-dependency-bloat-in-victoriametrics-docker-image-983fb5912b0d) without external dependencies.
* All the configuration is done via explicit command-line flags with reasonable defaults.
* All the data is stored in a single directory pointed by `-storageDataPath` flag.
* Easy and fast backups from [instant snapshots](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282)
to S3 or GCS with [vmbackup](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmbackup/README.md) / [vmrestore](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmrestore/README.md).
See [this article](https://medium.com/@valyala/speeding-up-backups-for-big-time-series-databases-533c1a927883) for more details.
* Storage is protected from corruption on unclean shutdown (i.e. OOM, hardware reset or `kill -9`) thanks to [the storage architecture](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282).
* Supports metrics' scraping, ingestion and [backfilling](#backfilling) via the following protocols:
* [Metrics from Prometheus exporters](https://github.com/prometheus/docs/blob/master/content/docs/instrumenting/exposition_formats.md#text-based-format)
such as [node_exporter](https://github.com/prometheus/node_exporter). See [these docs](#how-to-scrape-prometheus-exporters-such-as-node-exporter) for details.
* Ideally works with big amounts of time series data from Kubernetes, IoT sensors, connected cars, industrial telemetry, financial data and various Enterprise workloads.
* Has open source [cluster version](https://github.com/VictoriaMetrics/VictoriaMetrics/tree/cluster).
Just start VictoriaMetrics [executable](https://github.com/VictoriaMetrics/VictoriaMetrics/releases)
or [docker image](https://hub.docker.com/r/victoriametrics/victoria-metrics/) with the desired command-line flags.
The following command-line flags are used the most:
*`-storageDataPath` - path to data directory. VictoriaMetrics stores all the data in this directory. Default path is `victoria-metrics-data` in current working directory.
*`-retentionPeriod` - retention period in months for the data. Older data is automatically deleted. Default period is 1 month.
Each flag values can be set thru environment variables by following these rules:
* The `-envflag.enable` flag must be set
* Each `.` in flag names must be substituted by `_` (for example `-insert.maxQueueDuration <duration>` will translate to `insert_maxQueueDuration=<duration>`)
* For repeating flags, an alternative syntax can be used by joining the different values into one using `,` as separator (for example `-storageNode <nodeA> -storageNode <nodeB>` will translate to `storageNode=<nodeA>,<nodeB>`)
* It is possible setting prefix for environment vars with `-envflag.prefix`. For instance, if `-envflag.prefix=VM_`, then env vars must be prepended with `VM_`
say otherwise. It is safe skipping multiple versions during the upgrade unless [release notes](https://github.com/VictoriaMetrics/VictoriaMetrics/releases) say otherwise.
It is recommended performing regular upgrades to the latest version, since it may contain important bug fixes, performance optimizations or new features.
### How to scrape Prometheus exporters such as [node-exporter](https://github.com/prometheus/node_exporter)
VictoriaMetrics can be used as drop-in replacement for Prometheus for scraping targets configured in `prometheus.yml` config file according to [the specification](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#configuration-file).
Just set `-promscrape.config` command-line flag to the path to `prometheus.yml` config - and VictoriaMetrics should start scraping the configured targets.
Currently the following [scrape_config](https://prometheus.io/docs/prometheus/latest/configuration/configuration/#scrape_config) types are supported:
The file pointed by `-promscrape.config` may contain `%{ENV_VAR}` placeholders, which are substituted by the corresponding `ENV_VAR` environment variable values.
See also [vmagent](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmagent/README.md), which can be used as drop-in replacement for Prometheus.
### How to send data from InfluxDB-compatible agents such as [Telegraf](https://www.influxdata.com/time-series-platform/telegraf/)
Note that Influx line protocol expects [timestamps in *nanoseconds* by default](https://docs.influxdata.com/influxdb/v1.7/write_protocols/line_protocol_tutorial/#timestamp),
while VictoriaMetrics stores them with *milliseconds* precision.
*`time` - the corresponding CSV column at `<column_pos>` contains metric time. CSV line may contain either one or zero columns with time.
If CSV line has no time, then the current time is used. The time is applied to all the configured metrics.
The format of the time is configured via `<context>`. Supported time formats are:
*`unix_s` - unix timestamp in seconds.
*`unix_ms` - unix timestamp in milliseconds.
*`unix_ns` - unix timestamp in nanoseconds. Note that VictoriaMetrics rounds the timestamp to milliseconds.
*`rfc3339` - timestamp in [RFC3339](https://tools.ietf.org/html/rfc3339) format, i.e. `2006-01-02T15:04:05Z`.
*`custom:<layout>` - custom layout for the timestamp. The `<layout>` may contain arbitrary time layout according to [time.Parse rules in Go](https://golang.org/pkg/time/#Parse).
### How to import data in Prometheus exposition format
VictoriaMetrics accepts data in [Prometheus exposition format](https://github.com/prometheus/docs/blob/master/content/docs/instrumenting/exposition_formats.md#text-based-format)
via `/api/v1/import/prometheus` path. For example, the following line imports a single line in Prometheus exposition format into VictoriaMetrics:
```bash
curl -d 'foo{bar="baz"} 123' -X POST 'http://localhost:8428/api/v1/import/prometheus'
```
The following command may be used for verifying the imported data:
It can be overriden by passing unix timestamp in *milliseconds* via `timestamp` query arg. For example, `/api/v1/import/prometheus?timestamp=1594370496905`.
Additionally to unix timestamps and [RFC3339](https://www.ietf.org/rfc/rfc3339.txt) VictoriaMetrics accepts relative times in `time`, `start` and `end` query args.
For example, the following query would return data for the last 30 minutes: `/api/v1/query_range?start=-30m&query=...`.
By default, VictoriaMetrics returns time series for the last 5 minutes from /api/v1/series, while the Prometheus API defaults to all time. Use `start` and `end` to select a different time range.
*`/api/v1/series/count` - it returns the total number of time series in the database. Some notes:
* the handler scans all the inverted index, so it can be slow if the database contains tens of millions of time series;
* the handler may count [deleted time series](#how-to-delete-time-series) additionally to normal time series due to internal implementation restrictions;
VictoriaMetrics accepts the following additional query args at `/metrics/find` and `/metrics/expand`:
*`label` - for selecting arbitrary label values. By default `label=__name__`, i.e. metric names are selected.
*`delimiter` - for using different delimiters in metric name hierachy. For example, `/metrics/find?delimiter=_&query=node_*` would return all the metric name prefixes
that start with `node_`. By default `delimiter=.`.
helps to spin up VictoriaMetrics, [vmagent](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmagent/README.md) and Grafana with one command.
More details may be found [here](https://github.com/VictoriaMetrics/VictoriaMetrics/tree/master/deployment/docker#folder-contains-basic-images-and-tools-for-building-and-running-victoria-metrics-in-docker).
### Setting up service
Read [these instructions](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/43) on how to set up VictoriaMetrics as a service in your OS.
VictoriaMetrics can create [instant snapshots](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282)
for all the data stored under `-storageDataPath` directory.
Navigate to `http://<victoriametrics-addr>:8428/snapshot/create` in order to create an instant snapshot.
Send a request to `http://<victoriametrics-addr>:8428/api/v1/admin/tsdb/delete_series?match[]=<timeseries_selector_for_delete>`,
where `<timeseries_selector_for_delete>` may contain any [time series selector](https://prometheus.io/docs/prometheus/latest/querying/basics/#time-series-selectors)
for metrics to delete. After that all the time series matching the given selector are deleted. Storage space for
the deleted time series isn't freed instantly - it is freed during subsequent [background merges of data files](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282).
It is recommended verifying which metrics will be deleted with the call to `http://<victoria-metrics-addr>:8428/api/v1/series?match[]=<timeseries_selector_for_delete>`
before actually deleting the metrics. By default this query will only scan active series in the past 5 minutes, so you may need to
adjust `start` and `end` to a suitable range to achieve match hits.
VictoriaMetrics performs [data compations in background](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282)
in order to keep good performance characteristics when accepting new data. These compactions (merges) are performed independently on per-month partitions.
This means that compactions are stopped for per-month partitions if no new data is ingested into these partitions.
Sometimes it is necessary to trigger compactions for old partitions. For instance, in order to free up disk space occupied by [deleted time series](#how-to-delete-time-series).
In this case forced compaction may be initiated on the specified per-month partition by sending request to `/internal/force_merge?partition_prefix=YYYY_MM`,
where `YYYY_MM` is per-month partition name. For example, `http://victoriametrics:8428/internal/force_merge?partition_prefix=2020_08` would initiate forced
merge for August 2020 partition. The call to `/internal/force_merge` returns immediately, while the corresponding forced merges continues running in background.
Forced merges may require additional CPU, disk IO and storage space resources. It is unnecessary to run forced merge under normal conditions,
since VictoriaMetrics automatically performs [optimal merges in background](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282)
Send a request to `http://<victoriametrics-addr>:8428/api/v1/export?match[]=<timeseries_selector_for_export>`,
where `<timeseries_selector_for_export>` may contain any [time series selector](https://prometheus.io/docs/prometheus/latest/querying/basics/#time-series-selectors)
for metrics to export. Use `{__name__!=""}` selector for fetching all the time series.
The response would contain all the data for the selected time series in [JSON streaming format](https://en.wikipedia.org/wiki/JSON_streaming#Line-delimited_JSON).
*`/api/v1/import/prometheus` http POST handler, which accepts data in Prometheus exposition format. See [these docs](#how-to-import-data-in-prometheus-exposition-format) for details.
Each request to `/api/v1/import` can load up to a single vCPU core on VictoriaMetrics. Import speed can be improved by splitting the original file into smaller parts
and importing them concurrently. Note that the original file must be split on newlines.
*`replace_all`: replaces all the occurences of `regex` in the values of `source_labels` with the `replacement` and stores the result in the `target_label`.
*`labelmap_all`: replaces all the occurences of `regex` in all the label names with the `replacement`.
*`keep_if_equal`: keeps the entry if all label values from `source_labels` are equal.
*`drop_if_equal`: drops the entry if all the label values from `source_labels` are equal.
See also [relabeling in vmagent](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmagent/README.md#relabeling).
at `http://<victoriametrics-addr>:8428/federate?match[]=<timeseries_selector_for_federation>`.
Optional `start` and `end` args may be added to the request in order to scrape the last point for each selected time series on the `[start ... end]` interval.
`start` and `end` may contain either unix timestamp in seconds or [RFC3339](https://www.ietf.org/rfc/rfc3339.txt) values. By default, the last point
on the interval `[now - max_lookback ... now]` is scraped for each time series. The default value for `max_lookback` is `5m` (5 minutes), but it can be overridden.
For instance, `/federate?match[]=up&max_lookback=1h` would return last points on the `[now - 1h ... now]` interval. This may be useful for time series federation
with scrape intervals exceeding `5m`.
### Capacity planning
A rough estimation of the required resources for ingestion path:
* RAM size: less than 1KB per active time series. So, ~1GB of RAM is required for 1M active time series.
Time series is considered active if new data points have been added to it recently or if it has been recently queried.
The number of active time series may be obtained from `vm_cache_entries{type="storage/hour_metric_ids"}` metric
VictoriaMetrics stores various caches in RAM. Memory size for these caches may be limited with `-memory.allowedPercent` or `-memory.allowedBytes` flags.
* CPU cores: a CPU core per 300K inserted data points per second. So, ~4 CPU cores are required for processing
the insert stream of 1M data points per second. The ingestion rate may be lower for high cardinality data or for time series with high number of labels.
See [this article](https://medium.com/@valyala/insert-benchmarks-with-inch-influxdb-vs-victoriametrics-e31a41ae2893) for details.
If you see lower numbers per CPU core, then it is likely active time series info doesn't fit caches,
so you need more RAM for lowering CPU usage.
* Storage space: less than a byte per data point on average. So, ~260GB is required for storing a month-long insert stream
of 100K data points per second.
The actual storage size heavily depends on data randomness (entropy). Higher randomness means higher storage size requirements.
2) Pass addresses of these instances to [vmagent](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmagent/README.md) via `-remoteWrite.url` command-line flag:
It is recommended to use [vmagent](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmagent/README.md) instead of Prometheus for highly loaded setups.
on the same time series if they fall within the same discrete 60s bucket. The earliest data point will be kept. In the case of equal timestamps, an arbitrary data point will be kept.
Then set up [vmauth](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmauth/README.md) in front of VictoriaMetrics instances,
so it could route requests from particular user to VictoriaMetrics with the desired retention.
The same scheme could be implemented for multiple tenants in [VictoriaMetrics cluster](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/cluster/README.md).
* VictoriaMetrics has good compression for on-disk data. See [this article](https://medium.com/@valyala/victoriametrics-achieving-better-compression-for-time-series-data-than-gorilla-317bc1f95932)
Single-node VictoriaMetrics doesn't support multi-tenancy. Use [cluster version](https://github.com/VictoriaMetrics/VictoriaMetrics/tree/cluster) instead.
### Scalability and cluster version
Though single-node VictoriaMetrics cannot scale to multiple nodes, it is optimized for resource usage - storage size / bandwidth / IOPS, RAM, CPU.
This means that a single-node VictoriaMetrics may scale vertically and substitute a moderately sized cluster built with competing solutions
such as Thanos, Uber M3, InfluxDB or TimescaleDB. See [vertical scalability benchmarks](https://medium.com/@valyala/measuring-vertical-scalability-for-time-series-databases-in-google-cloud-92550d78d8ae).
So try single-node VictoriaMetrics at first and then [switch to cluster version](https://github.com/VictoriaMetrics/VictoriaMetrics/tree/cluster) if you still need
horizontally scalable long-term remote storage for really large Prometheus deployments.
[Contact us](mailto:info@victoriametrics.com) for paid support.
It is recommended using [vmalert](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmalert/README.md) for alerting.
Additionally, alerting can be set up with the following tools:
* With Prometheus - see [the corresponding docs](https://prometheus.io/docs/alerting/overview/).
* With Promxy - see [the corresponding docs](https://github.com/jacksontj/promxy/blob/master/README.md#how-do-i-use-alertingrecording-rules-in-promxy).
* With Grafana - see [the corresponding docs](https://grafana.com/docs/alerting/rules/).
Prefer authorizing all the incoming requests from untrusted networks with [vmauth](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmauth/README.md)
The only option is increasing the limit on [the number of open files in the OS](https://medium.com/@muhammadtriwibowo/set-permanently-ulimit-n-open-files-in-ubuntu-4d61064429a),
so Prometheus instances could establish more connections to VictoriaMetrics.
* The recommended filesystem is `ext4`, the recommended persistent storage is [persistent HDD-based disk on GCP](https://cloud.google.com/compute/docs/disks/#pdspecs),
since it is protected from hardware failures via internal replication and it can be [resized on the fly](https://cloud.google.com/compute/docs/disks/add-persistent-disk#resize_pd).
There are officials Grafana dashboards for [single-node VictoriaMetrics](https://grafana.com/dashboards/10229) and [clustered VictoriaMetrics](https://grafana.com/grafana/dashboards/11176).
* VictoriaMetrics requires free disk space for [merging data files to bigger ones](https://medium.com/@valyala/how-victoriametrics-makes-instant-snapshots-for-multi-terabyte-time-series-data-e1f3fb0e0282).
It may slow down when there is no enough free space left. So make sure `-storageDataPath` directory
* VictoriaMetrics ignores `NaN` and `Inf` values during data ingestion. See [this issue](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/752) for details.
Storage-level replication may be offloaded to durable persistent storage such as [Google Cloud disks](https://cloud.google.com/compute/docs/disks#pdspecs).
VictoriaMetrics supports backups via [vmbackup](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmbackup/README.md)
and [vmrestore](https://github.com/VictoriaMetrics/VictoriaMetrics/blob/master/app/vmrestore/README.md) tools.
We also provide provide `vmbackuper` tool for paid enterprise subscribers - see [this issue](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/466) for details.
* [Prometheus Oauth proxy](https://gitlab.com/optima_public/prometheus_oauth_proxy) - see [this article](https://medium.com/@richard.holly/powerful-saas-solution-for-detection-metrics-c67b9208d362) for details.
