Streaming and replication¶
Each netdata is able to replicate/mirror its database to another netdata, by streaming collected metrics, in real-time to it. This is quite different to data archiving to third party time-series databases.
When a netdata streams metrics to another netdata, the receiving one is able to perform everything a netdata performs:
- visualize them with a dashboard
- run health checks that trigger alarms and send alarm notifications
- archive metrics to a backend time-series database
netdata without a database or web API (headless collector)¶
Local netdata (
slave), without any database or alarms, collects metrics and sends them to
another netdata (
The node menu shows a list of all “databases streamed to” the master. Clicking one of those links allows the user to view the full dashboard of the
slave netdata. The URL has the form http://master-host:master-port/host/slave-host/.
Alarms for the
slave are served by the
In this mode the
slave is just a plain data collector. It spawns all external plugins, but instead
of maintaining a local database and accepting dashboard requests, it streams all metrics to the
master. The memory footprint is reduced significantly, to between 6 MiB and 40 MiB, depending on the enabled plugins. To reduce the memory usage as much as possible, refer to running netdata in embedded devices.
master can collect data for any number of
Local netdata (
slave), with a local database (and possibly alarms), collects metrics and
sends them to another netdata (
The user can use all the functions at both http://slave-ip:slave-port/ and http://master-host:master-port/host/slave-host/.
slave and the
master may have different data retention policies for the same metrics.
Alarms for the
slave are triggered by both the
slave and the
master (and actually
each can have different alarms configurations or have alarms disabled).
Take a note, that custom chart names, configured on the
slave, should be in the form
type.name to work correctly. The
master will truncate the
type part and substitute the original chart
type to store the name in the database.
Local netdata (
slave), with or without a database, collects metrics and sends them to another
proxy), which may or may not maintain a database, which forwards them to another
Alarms for the slave can be triggered by any of the involved hosts that maintains a database.
Any number of daisy chaining netdata servers are supported, each with or without a database and
with or without alarms for the
mix and match with backends¶
All nodes that maintain a database can also send their data to a backend database. This allows quite complex setups.
Bdo not maintain a database and stream metrics to netdata
C(live streaming functionality, i.e. this PR)
Cmaintains a database for
Cand archives all metrics to
graphitewith 10 second detail (backends functionality)
Calso streams data for
D, which also collects data from
Gfrom another DMZ (live streaming functionality, i.e. this PR)
Dis just a proxy, without a database, that streams all data to a remote site at netdata
Hmaintains a database for
Hand sends all data to
opentsdbwith 5 seconds detail (backends functionality)
- alarms are triggered by
Hfor all hosts
- users can use all the netdata that maintain a database to view metrics (i.e. at
Hall hosts can be viewed).
These are options that affect the operation of netdata in this area:
[global] memory mode = none | ram | save | map | dbengine
[global].memory mode = none disables the database at this host. This also disables health
monitoring (there cannot be health monitoring without a database).
[web] mode = none | static-threaded accept a streaming request every seconds = 0
[web].mode = none disables the API (netdata will not listen to any ports).
This also disables the registry (there cannot be a registry without an API).
accept a streaming request every seconds can be used to set a limit on how often a master Netdata server will accept streaming requests from the slaves. 0 sets no limit, 1 means maximum once every second. If this is set, you may see error log entries “… too busy to accept new streaming request. Will be allowed in X secs”.
[backend] enabled = yes | no type = graphite | opentsdb destination = IP:PORT ... update every = 10
[backend] configures data archiving to a backend (it archives all databases maintained on
A new file is introduced: stream.conf (to edit it on your system run
/etc/netdata/edit-config stream.conf). This file holds streaming configuration for both the
sending and the receiving netdata.
API keys are used to authorize the communication of a pair of sending-receiving netdata. Once the communication is authorized, the sending netdata can push metrics for any number of hosts.
You can generate an API key with the command
uuidgen. API keys are just random GUIDs.
You can use the same API key on all your netdata, or use a different API key for any pair of
options for the sending node¶
This is the section for the sending netdata. On the receiving node,
[stream].enabled can be
If it is
yes, the receiving node will also stream the metrics to another node (i.e. it will be
[stream] enabled = yes | no destination = IP:PORT[:SSL] ... api key = XXXXXXXXXXX
This is an overview of how these options can be combined:
|proxy with db||not
For the options to encrypt the data stream between the slave and the master, refer to securing the communication
options for the receiving node¶
stream.conf looks like this:
# replace API_KEY with your uuidgen generated GUID [API_KEY] enabled = yes default history = 3600 default memory mode = save health enabled by default = auto allow from = *
You can add many such sections, one for each API key. The above are used as default values for all hosts pushed with this API key.
You can also add sections like this:
# replace MACHINE_GUID with the slave /var/lib/netdata/registry/netdata.public.unique.id [MACHINE_GUID] enabled = yes history = 3600 memory mode = save health enabled = yes allow from = *
The above is the receiver configuration of a single host, at the receiver end.
the unique id the netdata generating the metrics (i.e. the netdata that originally collects
/var/lib/netdata/registry/netdata.unique.id). So, metrics for netdata
A that pass through
any number of other netdata, will have the same
You can also use
default memory mode = dbengine for an API key or
memory mode = dbengine for
a single host. The additional
page cache size and
dbengine disk space configuration options
are inherited from the global netdata configuration.
allow from settings are netdata simple patterns: string matches
* as wildcard (any number of times) and a
! prefix for a negative match.
allow from = !10.1.2.3 10.* will allow all IPs in
10.1.2.3. The order is
important: left to right, the first positive or negative match is used.
allow from is available in netdata v1.9+
slave is trying to push metrics to a
proxy, it logs entries like these:
2017-02-25 01:57:44: netdata: ERROR: Failed to connect to '10.11.12.1', port '19999' (errno 111, Connection refused) 2017-02-25 01:57:44: netdata: ERROR: STREAM costa-pc [send to 10.11.12.1:19999]: failed to connect 2017-02-25 01:58:04: netdata: INFO : STREAM costa-pc [send to 10.11.12.1:19999]: initializing communication... 2017-02-25 01:58:04: netdata: INFO : STREAM costa-pc [send to 10.11.12.1:19999]: waiting response from remote netdata... 2017-02-25 01:58:14: netdata: INFO : STREAM costa-pc [send to 10.11.12.1:19999]: established communication - sending metrics... 2017-02-25 01:58:14: netdata: ERROR: STREAM costa-pc [send]: discarding 1900 bytes of metrics already in the buffer. 2017-02-25 01:58:14: netdata: INFO : STREAM costa-pc [send]: ready - sending metrics...
The receiving end (
master) logs entries like these:
2017-02-25 01:58:04: netdata: INFO : STREAM [receive from [10.11.12.11]:33554]: new client connection. 2017-02-25 01:58:04: netdata: INFO : STREAM costa-pc [10.11.12.11]:33554: receive thread created (task id 7698) 2017-02-25 01:58:14: netdata: INFO : Host 'costa-pc' with guid '12345678-b5a6-11e6-8a50-00508db7e9c9' initialized, os: linux, update every: 1, memory mode: ram, history entries: 3600, streaming: disabled, health: enabled, cache_dir: '/var/cache/netdata/12345678-b5a6-11e6-8a50-00508db7e9c9', varlib_dir: '/var/lib/netdata/12345678-b5a6-11e6-8a50-00508db7e9c9', health_log: '/var/lib/netdata/12345678-b5a6-11e6-8a50-00508db7e9c9/health/health-log.db', alarms default handler: '/usr/libexec/netdata/plugins.d/alarm-notify.sh', alarms default recipient: 'root' 2017-02-25 01:58:14: netdata: INFO : STREAM costa-pc [receive from [10.11.12.11]:33554]: initializing communication... 2017-02-25 01:58:14: netdata: INFO : STREAM costa-pc [receive from [10.11.12.11]:33554]: receiving metrics...
For netdata v1.9+, streaming can also be monitored via
Securing streaming communications¶
Netdata does not activate TLS encryption by default. To encrypt streaming connections, you first need to enable TLS support on the master. With encryption enabled on the receiving side, you need to instruct the slave to use TLS/SSL as well. On the slave’s
stream.conf, configure the destination as follows:
[stream] destination = host:port:SSL
SSL appended to the end of the destination tells the slave that connections must be encrypted.
Differences in TLS and SSL terminology
While Netdata uses Transport Layer Security (TLS) 1.2 to encrypt communications rather than the obsolete SSL protocol, it’s still common practice to refer to encrypted web connections as
SSL. Many vendors, like Nginx and even Netdata itself, use
SSL in configuration files, whereas documentation will always refer to encrypted communications as
When TLS/SSL is enabled on the slave, the default behavior will be to not connect with the master unless the server’s certificate can be verified via the default chain. In case you want to avoid this check, add the following to the slave’s
[stream] ssl skip certificate verification = yes
With the introduction of TLS/SSL, the master-slave communication behaves as shown in the table below, depending on the following configurations:
- Master TLS (Yes/No): Whether the
- Master port TLS (-/force/optional): Depends on whether the
bind tocontains a
^SSL=optionaldirective on the port(s) used for streaming.
- Slave TLS (Yes/No): Whether the destination in the slave’s
:SSLat the end.
- Slave TLS Verification (yes/no): Value of the slave’s
ssl skip certificate verificationparameter (default is no).
|Master TLS enabled||Master port SSL||Slave TLS||Slave SSL Ver.||Behavior|
|No||-||No||no||Legacy behavior. The master-slave stream is unencrypted.|
|Yes||force||No||no||The master rejects the slave connection.|
|Yes||-/optional||No||no||The master-slave stream is unencrypted (expected situation for legacy slaves and newer masters)|
|Yes||-/force/optional||Yes||no||The master-slave stream is encrypted, provided that the master has a valid TLS/SSL certificate. Otherwise, the slave refuses to connect.|
|Yes||-/force/optional||Yes||yes||The master-slave stream is encrypted.|
Viewing remote host dashboards, using mirrored databases¶
On any receiving netdata, that maintains remote databases and has its web server enabled, The node menu will include a list of the mirrored databases.
Selecting any of these, the server will offer a dashboard using the mirrored metrics.
Monitoring ephemeral nodes¶
Auto-scaling is probably the most trendy service deployment strategy these days.
Auto-scaling detects the need for additional resources and boots VMs on demand, based on a template. Soon after they start running the applications, a load balancer starts distributing traffic to them, allowing the service to grow horizontally to the scale needed to handle the load. When demands falls, auto-scaling starts shutting down VMs that are no longer needed.
What a fantastic feature for controlling infrastructure costs! Pay only for what you need for the time you need it!
In auto-scaling, all servers are ephemeral, they live for just a few hours. Every VM is a brand new instance of the application, that was automatically created based on a template.
So, how can we monitor them? How can we be sure that everything is working as expected on all of them?
The netdata way¶
We recently made a significant improvement at the core of netdata to support monitoring such setups.
Following the netdata way of monitoring, we wanted:
- real-time performance monitoring, collecting thousands of metrics per server per second, visualized in interactive, automatically created dashboards.
- real-time alarms, for all nodes.
- zero configuration, all ephemeral servers should have exactly the same configuration, and nothing should be configured at any system for each of the ephemeral nodes. We shouldn’t care if 10 or 100 servers are spawned to handle the load.
- self-cleanup, so that nothing needs to be done for cleaning up the monitoring infrastructure from the hundreds of nodes that may have been monitored through time.
How it works¶
All monitoring solutions, including netdata, work like this:
collect metrics, from the system and the running applications
store metrics, in a time-series database
examine metricsperiodically, for triggering alarms and sending alarm notifications
visualize metrics, so that users can see what exactly is happening
netdata used to be self-contained, so that all these functions were handled entirely by each server. The changes we made, allow each netdata to be configured independently for each function. So, each netdata can now act as:
self contained system, much like it used to be.
data collector, that collects metrics from a host and pushes them to another netdata (with or without a local database and alarms).
proxy, that receives metrics from other hosts and pushes them immediately to other netdata servers. netdata proxies can also be
store and forward proxiesmeaning that they are able to maintain a local database for all metrics passing through them (with or without alarms).
time-series databasenode, where data are kept, alarms are run and queries are served to visualise the metrics.
Configuring an auto-scaling setup¶
You need a netdata
master. This node should not be ephemeral. It will be the node where all ephemeral nodes (let’s call them
slaves) will be sending their metrics.
The master will need to authorize the slaves for accepting their metrics. This is done with an API key.
API keys are just random GUIDs. Use the Linux command
uuidgen to generate one. You can use the same API key for all your
slaves, or you can configure one API for each of them. This is entirely your decision.
We suggest to use the same API key for each ephemeral node template you have, so that all replicas of the same ephemeral node will have exactly the same configuration.
I will use this API_KEY:
11111111-2222-3333-4444-555555555555. Replace it with your own.
On the master, edit
/etc/netdata/stream.conf (to edit it on your system run
/etc/netdata/edit-config stream.conf) and set these:
[11111111-2222-3333-4444-555555555555] # enable/disable this API key enabled = yes # one hour of data for each of the slaves default history = 3600 # do not save slave metrics on disk default memory = ram # alarms checks, only while the slave is connected health enabled by default = auto
stream.confon master, to enable receiving metrics from slaves using the API key.
If you used many API keys, you can add one such section for each API key.
When done, restart netdata on the
master node. It is now ready to receive metrics.
health enabled by default = auto will still trigger
last_collected alarms, if a connected slave does not exit gracefully. If the netdata running on the slave is
stopped, it will close the connection to the master, ensuring that no
last_collected alarms are triggered. For example, a proper container restart would first terminate
the netdata process, but a system power issue would leave the connection open on the master side. In the second case, you will still receive alarms.
On each of the slaves, edit
/etc/netdata/stream.conf (to edit it on your system run
/etc/netdata/edit-config stream.conf) and set these:
[stream] # stream metrics to another netdata enabled = yes # the IP and PORT of the master destination = 10.11.12.13:19999 # the API key to use api key = 11111111-2222-3333-4444-555555555555
stream.confon slaves, to enable pushing metrics to master at
Using just the above configuration, the
slaves will be pushing their metrics to the
master netdata, but they will still maintain a local database of the metrics and run health checks. To disable them, edit
/etc/netdata/netdata.conf and set:
[global] # disable the local database memory mode = none [health] # disable health checks enabled = no
netdata.confconfiguration on slaves, to disable the local database and health checks.
Keep in mind that setting
memory mode = none will also force
[health].enabled = no (health checks require access to a local database). But you can keep the database and disable health checks if you need to. You are however sending all the metrics to the master server, which can handle the health checking (
[health].enabled = yes)
netdata unique id¶
/var/lib/netdata/registry/netdata.public.unique.id contains a random GUID that uniquely identifies each netdata. This file is automatically generated, by netdata, the first time it is started and remains unaltered forever.
If you are building an image to be used for automated provisioning of autoscaled VMs, it important to delete that file from the image, so that each instance of your image will generate its own.
Troubleshooting metrics streaming¶
Both the sender and the receiver of metrics log information at
On both master and slave do this:
tail -f /var/log/netdata/error.log | grep STREAM
If the slave manages to connect to the master you will see something like (on the master):
2017-03-09 09:38:52: netdata: INFO : STREAM [receive from [10.11.12.86]:38564]: new client connection. 2017-03-09 09:38:52: netdata: INFO : STREAM xxx [10.11.12.86]:38564: receive thread created (task id 27721) 2017-03-09 09:38:52: netdata: INFO : STREAM xxx [receive from [10.11.12.86]:38564]: client willing to stream metrics for host 'xxx' with machine_guid '1234567-1976-11e6-ae19-7cdd9077342a': update every = 1, history = 3600, memory mode = ram, health auto 2017-03-09 09:38:52: netdata: INFO : STREAM xxx [receive from [10.11.12.86]:38564]: initializing communication... 2017-03-09 09:38:52: netdata: INFO : STREAM xxx [receive from [10.11.12.86]:38564]: receiving metrics...
and something like this on the slave:
2017-03-09 09:38:28: netdata: INFO : STREAM xxx [send to box:19999]: connecting... 2017-03-09 09:38:28: netdata: INFO : STREAM xxx [send to box:19999]: initializing communication... 2017-03-09 09:38:28: netdata: INFO : STREAM xxx [send to box:19999]: waiting response from remote netdata... 2017-03-09 09:38:28: netdata: INFO : STREAM xxx [send to box:19999]: established communication - sending metrics...
Archiving to a time-series database¶
master netdata node can also archive metrics, for all
slaves, to a time-series database. At the time of this writing, netdata supports:
- json document DBs
- all the compatibles to the above (e.g. kairosdb, influxdb, etc)
Check the netdata backends documentation for configuring this.
This is how such a solution will work:
An advanced setup¶
netdata also supports
proxies with and without a local database, and data retention can be different between all nodes.
This means a setup like the following is also possible:
A proxy is a netdata that is receiving metrics from a netdata, and streams them to another netdata.
netdata proxies may or may not maintain a database for the metrics passing through them. When they maintain a database, they can also run health checks (alarms and notifications) for the remote host that is streaming the metrics.
To configure a proxy, configure it as a receiving and a sending netdata at the same time, using stream.conf.
The sending side of a netdata proxy, connects and disconnects to the final destination of the metrics, following the same pattern of the receiving side.
For a practical example see Monitoring ephemeral nodes.