⚠️ The backends system is now deprecated in favor of the exporting engine.
Netdata supports backends for archiving the metrics, or providing long term dashboards, using Grafana or other tools, like this:
Since Netdata collects thousands of metrics per server per second, which would easily congest any backend server when several Netdata servers are sending data to it, Netdata allows sending metrics at a lower frequency, by resampling them.
So, although Netdata collects metrics every second, it can send to the backend servers averages or sums every X seconds (though, it can send them per second if you need it to).
plaintext interface, used by Graphite, InfluxDB, KairosDB, Blueflood, ElasticSearch via logstash tcp input and the graphite codec, etc)
metrics are sent to the backend server as
prefixis configured below,
hostnameis the hostname of the machine (can also be configured).
telnet or HTTP interfaces, used by OpenTSDB, InfluxDB, KairosDB, etc)
metrics are sent to opentsdb as
json document DBs
metrics are sent to a document db,
prometheus is described at prometheus page since it pulls data from Netdata.
prometheus remote write (a binary snappy-compressed protocol buffer encoding over HTTP used by Elasticsearch, Gnocchi, Graphite, InfluxDB, Kafka, OpenTSDB, PostgreSQL/TimescaleDB, Splunk, VictoriaMetrics, and a lot of other storage providers)
TimescaleDB via community-built connector that takes JSON streams from a Netdata client and writes them to a TimescaleDB table.
AWS Kinesis Data Streams
metrics are sent to the service in
metrics are sent to the database in
Only one backend may be active at a time.
Netdata can filter metrics (at the chart level), to send only a subset of the collected metrics.
Netdata supports three modes of operation for all backends:
as-collectedsends to backends the metrics as they are collected, in the units they are collected. So, counters are sent as counters and gauges are sent as gauges, much like all data collectors do. For example, to calculate CPU utilization in this format, you need to know how to convert kernel ticks to percentage.
averagesends to backends normalized metrics from the Netdata database. In this mode, all metrics are sent as gauges, in the units Netdata uses. This abstracts data collection and simplifies visualization, but you will not be able to copy and paste queries from other sources to convert units. For example, CPU utilization percentage is calculated by Netdata, so Netdata will convert ticks to percentage and send the average percentage to the backend.
volume: the sum of the interpolated values shown on the Netdata graphs is sent to the backend. So, if Netdata is configured to send data to the backend every 10 seconds, the sum of the 10 values shown on the Netdata charts will be used.
Time-series databases suggest to collect the raw values (
as-collected). If you plan to invest on building your monitoring around a time-series database and you already know (or you will invest in learning) how to convert units and normalize the metrics in Grafana or other visualization tools, we suggest to use
If, on the other hand, you just need long term archiving of Netdata metrics and you plan to mainly work with Netdata, we suggest to use
average. It decouples visualization from data collection, so it will generally be a lot simpler. Furthermore, if you use
average, the charts shown in the back-end will match exactly what you see in Netdata, which is not necessarily true for the other modes of operation.
This code is smart enough, not to slow down Netdata, independently of the speed of the backend server.
/etc/netdata/netdata.conf you should have something like this (if not download the latest version of
from your Netdata):
enabled = yes | no, enables or disables sending data to a backend
type = graphite | opentsdb:telnet | opentsdb:http | opentsdb:https | json | kinesis | mongodb, selects the backend type
destination = host1 host2 host3 ..., accepts a space separated list of hostnames, IPs (IPv4 and IPv6) and ports to connect to. Netdata will use the first available to send the metrics.
The format of each item in this list, is:
tcpis the default and only supported by the current backends.
[XX:XX...XX:XX](IPv6). For IPv6 you can to enclose the IP in
to separate it from the port.
PORTcan be a number of a service name. If omitted, the default port for the backend will be used (graphite = 2003, opentsdb = 4242).
Example IPv6 and IPv4 together:
When multiple servers are defined, Netdata will try the next one when the first one fails. This allows you to load-balance different servers: give your backend servers in different order on each Netdata.
Netdata also ships
nc-backend.sh, a script that can be used as a fallback backend to save the
metrics to disk and push them to the time-series database when it becomes available again. It can also be used to
monitor / trace / debug the metrics Netdata generates.
For kinesis backend
destination should be set to an AWS region (for example,
The MongoDB backend doesn't use the
destination option for its configuration. It uses the
configuration file instead.
data source = as collected, or
data source = average, or
data source = sum, selects the kind of data that will be sent to the backend.
hostname = my-name, is the hostname to be used for sending data to the backend server. By default this is
prefix = Netdata, is the prefix to add to all metrics.
update every = 10, is the number of seconds between sending data to the backend. Netdata will add some randomness to this number, to prevent stressing the backend server when many Netdata servers send data to the same backend. This randomness does not affect the quality of the data, only the time they are sent.
buffer on failures = 10, is the number of iterations (each iteration is
[backend].update everyseconds) to buffer data, when the backend is not available. If the backend fails to receive the data after that many failures, data loss on the backend is expected (Netdata will also log it).
timeout ms = 20000, is the timeout in milliseconds to wait for the backend server to process the data. By default this is
2 * update_every * 1000.
send hosts matching = localhost *includes one or more space separated patterns, using
*as wildcard (any number of times within each pattern). The patterns are checked against the hostname (the localhost is always checked as
localhost), allowing us to filter which hosts will be sent to the backend when this Netdata is a central Netdata aggregating multiple hosts. A pattern starting with
!gives a negative match. So to match all hosts named
*db*except hosts containing
!*child* *db*(so, the order is important: the first pattern matching the hostname will be used - positive or negative).
send charts matching = *includes one or more space separated patterns, using
*as wildcard (any number of times within each pattern). The patterns are checked against both chart id and chart name. A pattern starting with
!gives a negative match. So to match all charts named
apps.*except charts ending in
!*reads apps.*(so, the order is important: the first pattern matching the chart id or the chart name will be used - positive or negative).
send names instead of ids = yes | nocontrols the metric names Netdata should send to backend. Netdata supports names and IDs for charts and dimensions. Usually IDs are unique identifiers as read by the system and names are human friendly labels (also unique). Most charts and metrics have the same ID and name, but in several cases they are different: disks with device-mapper, interrupts, QoS classes, statsd synthetic charts, etc.
host tags = list of TAG=VALUEdefines tags that should be appended on all metrics for the given host. These are currently only sent to graphite, json, opentsdb and prometheus. Please use the appropriate format for each time-series db. For example opentsdb likes them like
TAG1=VALUE1 TAG2=VALUE2, but prometheus like
tag1="value1", tag2="value2". Host tags are mirrored with database replication (streaming of metrics between Netdata servers).
Starting from Netdata v1.20 the host tags are parsed in accordance with a configured backend type and stored as host labels so that they can be reused in API responses and exporting connectors. The parsing is supported for graphite, json, opentsdb, and prometheus (default) backend types. You can check how the host tags were parsed using the /api/v1/info API call.
Netdata provides 5 charts:
Buffered metrics, the number of metrics Netdata added to the buffer for dispatching them to the backend server.
Buffered data size, the amount of data (in KB) Netdata added the buffer.
Backend latency, the time the backend server needed to process the data Netdata sent. If there was a re-connection involved, this includes the connection time.(this chart has been removed, because it only measures the time Netdata needs to give the data to the O/S - since the backend servers do not ack the reception, Netdata does not have any means to measure this properly).
Backend operations, the number of operations performed by Netdata.
Backend thread CPU usage, the CPU resources consumed by the Netdata thread, that is responsible for sending the metrics to the backend server.
Netdata adds 4 alarms:
backend_last_buffering, number of seconds since the last successful buffering of backend data
backend_metrics_sent, percentage of metrics sent to the backend server
backend_metrics_lost, number of metrics lost due to repeating failures to contact the backend server
(this was misleading and has been removed).
backend_slow, the percentage of time between iterations needed by the backend time to process the data sent by Netdata