a0a473125b
After moving libnetwork to this repo, we need to update all the import paths for libnetwork to point to docker/docker/libnetwork instead of docker/libnetwork. This change implements that. Signed-off-by: Brian Goff <cpuguy83@gmail.com> |
||
---|---|---|
.. | ||
daemon.json | ||
Dockerfile.client | ||
Dockerfile.dind | ||
main.go | ||
README.md |
description | keywords | title |
---|---|---|
Learn to use the built-in network debugger to debug overlay networking problems | network, troubleshooting, debug | Debug overlay or swarm networking issues |
WARNING This tool can change the internal state of the libnetwork API, be really mindful on its use and read carefully the following guide. Improper use of it will damage or permanently destroy the network configuration.
Docker CE 17.12 and higher introduce a network debugging tool designed to help debug issues with overlay networks and swarm services running on Linux hosts. When enabled, a network diagnostic server listens on the specified port and provides diagnostic information. The network debugging tool should only be started to debug specific issues, and should not be left running all the time.
Information about networks is stored in the database, which can be examined using the API. Currently the database contains information about the overlay network as well as the service discovery data.
The Docker API exposes endpoints to query and control the network debugging tool. CLI integration is provided as a preview, but the implementation is not yet considered stable and commands and options may change without notice.
The tool is available into 2 forms:
- client only: dockereng/network-diagnostic:onlyclient
- docker in docker version: dockereng/network-diagnostic:17.12-dind The latter allows to use the tool with a cluster running an engine older than 17.12
Enable the diagnostic server
The tool currently only works on Docker hosts running on Linux. To enable it on a node follow the step below.
-
Set the
network-diagnostic-port
to a port which is free on the Docker host, in the/etc/docker/daemon.json
configuration file.“network-diagnostic-port”: <port>
-
Get the process ID (PID) of the
dockerd
process. It is the second field in the output, and is typically a number from 2 to 6 digits long.$ ps aux |grep dockerd | grep -v grep
-
Reload the Docker configuration without restarting Docker, by sending the
HUP
signal to the PID you found in the previous step.kill -HUP <pid-of-dockerd>
If systemd is used the command systemctl reload docker
will be enough
A message like the following will appear in the Docker host logs:
Starting the diagnostic server listening on <port> for commands
Disable the diagnostic tool
Repeat these steps for each node participating in the swarm.
-
Remove the
network-diagnostic-port
key from the/etc/docker/daemon.json
configuration file. -
Get the process ID (PID) of the
dockerd
process. It is the second field in the output, and is typically a number from 2 to 6 digits long.$ ps aux |grep dockerd | grep -v grep
-
Reload the Docker configuration without restarting Docker, by sending the
HUP
signal to the PID you found in the previous step.kill -HUP <pid-of-dockerd>
A message like the following will appear in the Docker host logs:
Disabling the diagnostic server
Access the diagnostic tool's API
The network diagnostic tool exposes its own RESTful API. To access the API, send a HTTP request to the port where the tool is listening. The following commands assume the tool is listening on port 2000.
Examples are not given for every endpoint.
Get help
$ curl localhost:2000/help
OK
/updateentry
/getentry
/gettable
/leavenetwork
/createentry
/help
/clusterpeers
/ready
/joinnetwork
/deleteentry
/networkpeers
/
/join
Join or leave the network database cluster
$ curl localhost:2000/join?members=ip1,ip2,...
$ curl localhost:2000/leave?members=ip1,ip2,...
ip1
, ip2
, ... are the swarm node ips (usually one is enough)
Join or leave a network
$ curl localhost:2000/joinnetwork?nid=<network id>
$ curl localhost:2000/leavenetwork?nid=<network id>
network id
can be retrieved on the manager with docker network ls --no-trunc
and has
to be the full length identifier
List cluster peers
$ curl localhost:2000/clusterpeers
List nodes connected to a given network
$ curl localhost:2000/networkpeers?nid=<network id>
network id
can be retrieved on the manager with docker network ls --no-trunc
and has
to be the full length identifier
Dump database tables
The tables are called endpoint_table
and overlay_peer_table
.
The overlay_peer_table
contains all the overlay forwarding information
The endpoint_table
contains all the service discovery information
$ curl localhost:2000/gettable?nid=<network id>&tname=<table name>
Interact with a specific database table
The tables are called endpoint_table
and overlay_peer_table
.
$ curl localhost:2000/<method>?nid=<network id>&tname=<table name>&key=<key>[&value=<value>]
Note: operations on tables have node ownership, this means that are going to remain persistent till the node that inserted them is part of the cluster
Access the diagnostic tool's CLI
The CLI is provided as a preview and is not yet stable. Commands or options may change at any time.
The CLI executable is called diagnosticClient
and is made available using a
standalone container.
docker run --net host dockereng/network-diagnostic:onlyclient -v -net <full network id> -t sd
The following flags are supported:
Flag | Description |
---|---|
-t | Table one of sd or overlay . |
-ip | The IP address to query. Defaults to 127.0.0.1. |
-net | The target network ID. |
-port | The target port. (default port is 2000) |
-a | Join/leave network |
-v | Enable verbose output. |
NOTE
By default the tool won't try to join the network. This is following the intent to not change
the state on which the node is when the diagnostic client is run. This means that it is safe
to run the diagnosticClient against a running daemon because it will just dump the current state.
When using instead the diagnosticClient in the containerized version the flag -a
MUST be passed
to avoid retrieving empty results. On the other side using the -a
flag against a loaded daemon
will have the undesirable side effect to leave the network and so cutting down the data path for
that daemon.
Container version of the diagnostic tool
The CLI is provided as a container with a 17.12 engine that needs to run using privileged mode.
NOTE
Remember that table operations have ownership, so any create entry
will be persistent till
the diagnostic container is part of the swarm.
-
Make sure that the node where the diagnostic client will run is not part of the swarm, if so do
docker swarm leave -f
-
To run the container, use a command like the following:
$ docker container run --name net-diagnostic -d --privileged --network host dockereng/network-diagnostic:17.12-dind
-
Connect to the container using
docker exec -it <container-ID> sh
, and start the server using the following command:$ kill -HUP 1
-
Join the diagnostic container to the swarm, then run the diagnostic CLI within the container.
$ ./diagnosticClient <flags>...
-
When finished debugging, leave the swarm and stop the container.
Examples
The following commands dump the service discovery table and verify node ownership.
NOTE
Remember to use the full network ID, you can easily find that with docker network ls --no-trunc
Service discovery and load balancer:
$ diagnostiClient -t sd -v -net n8a8ie6tb3wr2e260vxj8ncy4 -a
Overlay network:
$ diagnostiClient -port 2001 -t overlay -v -net n8a8ie6tb3wr2e260vxj8ncy4 -a