Awesome
decompose
Reverse-engineering tool for docker environments.
Takes all network connections from your docker containers and exports them as:
- graphviz dot
- structurizr dsl
- compose yaml
- plant uml
- pseudographical tree
- json stream
- statistics - nodes, connections and listen ports counts
- CSV with columns:
name
,listen
andoutbounds
rationale
I was in need for a tool to visualize and inspect big (more than 470 containers) dockerized legacy system without any schemes and having a bare minimum of documentation
analogs
Closest analogs, i can find, that not suit my needs very well:
- Red5d/docker-autocompose - produces only
compose yaml
- justone/dockviz - produces only
dot
, links and ports are taken from compose configuration (links
andports
sections) directly, therefore can miss some of them - LeoVerto/docker-network-graph - very same as above, in python
- weaveworks/scope - deprecated, no cli
features
- os-independent, it uses different strategies to get container connections:
- running on linux as root is the fastest way and it will work with all types of containers (even
scratch
-based) as it usensenter
- running as non-root or on non-linux OS will attempt to run
netsat
inside container, if this fails (i.e. for missingnetstat
binary), no connections for such container will be gathered
- running on linux as root is the fastest way and it will work with all types of containers (even
- single-binary, static-compiled unix-way
cli
(all output goes to stdout, progress information to stderr) - produces detailed connections graph with ports
- save
json
stream once and process it later in any way you want - all output formats are sorted, thus can be placed to any
vcs
to observe changes - fast, scans ~470 containers with ~4000 connections in around 5 sec
- auto-clusterization based on graph topology
- deep inspection mode, in wich connections between procesess inside containers, also collected and shown
- unix-sockets connections
- over 95% test-coverage
known limitations
- only established and listen connections are listed (but script like snapshots.sh can beat this)
composer-yaml
is not intended to be working out from the box, it can lack some of crucial information (even in-full
mode), or may contains cycles between nodes (removinglinks
section in services may help), its main purpose is for system overview- gephi fails to load edges from resulting graphviz, this can be fixed by any auto-replacement
tool:
sed -i 's/->/ -> /g' myfile.dot
- unix-sockets works only in root mode on linux, this process involves inode matching to find correct connections
installation
- binaries / deb / rpm for Linux, FreeBSD, macOS and Windows
- docker image
usage
decompose [flags]
-cluster string
json file with clusterization rules, or auto:<similarity> for auto-clustering, similarity is float in (0.0, 1.0] range
-compress
compress graph
-deep
process-based introspection
-follow string
follow only this container by name(s), comma-separated or from @file
-format string
output format: csv, dot, json, puml, sdsl, stat, tree, yaml (default "json")
-help
show this help
-load value
load json stream, can be used multiple times
-local
skip external hosts
-meta string
json file with metadata for enrichment
-no-loops
remove connection loops (node to itself) from output
-no-orphans
remove orphaned (not connected) nodes from output
-out string
output: filename or "-" for stdout (default "-")
-proto string
protocol to scan: tcp,udp,unix or all (default "all")
-silent
suppress progress messages in stderr
-skip-env string
environment variables name(s) to skip from output, case-independent, comma-separated
-version
show version
environment variables:
DOCKER_HOST
- connection uriDOCKER_CERT_PATH
- directory path containing key.pem, cert.pm and ca.pemDOCKER_TLS_VERIFY
- enable client TLS verificationIN_DOCKER_PROC_ROOT
- for in-docker scenario - root for host-mounted /proc
json stream format
type Item struct {
Name string `json:"name"` // container name
IsExternal bool `json:"is_external"` // this host is external
Image *string `json:"image,omitempty"` // docker image (if any)
Container struct{
Cmd []string `json:"cmd"`
Env []string `json:"env"`
Labels map[string]string `json:"labels"`
} `json:"container"` // container info
Listen map[string][]{
Kind string `json:"kind"` // tcp / udp / unix
Value string `json:"value"`
Local bool `json:"local"` // bound to loopback
} `json:"listen"` // ports with process names
Networks []string `json:"networks"` // network names
Tags []string `json:"tags"` // tags, if meta presents
Volumes []*struct{
Type string `json:"type"`
Src string `json:"src"`
Dst string `json:"dst"`
} `json:"volumes"` // volumes info, only when '-full'
Connected map[string][]string `json:"connected"` // name -> ports slice
}
Single node example with full info and metadata filled:
{
"name": "foo-1",
"is_external": false,
"image": "repo/foo:latest",
"container": {
"cmd": [
"foo",
"-foo-arg"
],
"env": [
"FOO=1"
],
"labels": {}
},
"listen": {"foo": [
{"kind": "tcp", "value": "80"}
]},
"networks": ["test-net"],
"tags": ["some"],
"volumes": [
{
"type": "volume",
"src": "/var/lib/docker/volumes/foo_1/_data",
"dst": "/data"
},
{
"type": "bind",
"src": "/path/to/foo.conf",
"dst": "/etc/foo.conf"
}
],
"connected": {
"bar-1": [
{"src": "foo", "dst": "[remote]", "port": {"kind": "tcp", "value": "443"}}
]
}
}
See stream.json for simple stream example.
metadata format
To enrich output with detailed descriptions, you can provide additional json
file, with metadata i.e.:
{
"foo": {
"info": "info for foo",
"docs": "https://acme.corp/docs/foo",
"repo": "https://git.acme.corp/foo",
"tags": ["some"]
},
"bar": {
"info": "info for bar",
"tags": ["other", "not-foo"]
}
}
Using this file decompose
can enrich output with info and additional tags, for every container that match by name with
one of provided keys, like foo-1
or bar1
for this example.
See csv2meta.py for example how to create such json
fom csv, and
meta.json for metadata sample.
clusterization
with rules
You can join your services into clusters
by flexible rules, in dot
, structurizr
and stat
output formats.
Example json
(order matters):
[
{
"name": "cluster-name",
"weight": 1,
"if": "<expression>"
},
...
]
Weight can be omitted, if not specified it equals 1
.
Where <expression>
is expr dsl, having env object node
with follownig
fields:
type Node struct {
Listen PortMatcher // port matcher with two methods: `HasAny(...string) bool` and `Has(...string) bool`
Name string // container name
Image string // container image
Cmd string // container cmd
Args []string // container args
Tags []string // tags, if meta present
IsExternal bool // external flag
}
See: cluster.json for detailed example.
automatic
Decompose provides automatic clusterization option, use -cluster auto:<similarity>
to try it out, similarity
is
a float in (0.0, 1.0]
range, representing how much similar ports nodes must have to be placed in same cluster
(1.0
- must have all ports equal).
examples
Save full json stream:
sudo decompose > nodes-1.json
Get dot
file:
decompose -format dot > connections.dot
Get tcp and udp connections as dot
:
decompose -proto tcp,udp -format dot > tcp.dot
Merge graphs from json streams, filter by protocol, skip remote hosts and save as dot
:
decompose -local -proto tcp -load "nodes-*.json" -format dot > graph-merged.dot
Load json stream, enrich and save as structurizr dsl
:
decompose -load nodes-1.json -meta metadata.json -format sdsl > workspace.dsl
Save auto-clustered graph, with similarity factor 0.6
as structurizr dsl
:
decompose -cluster auto:0.6 -format sdsl > workspace.dsl
example result
Scheme taken from redis-cluster:
it may be too heavy to display it with
browser, use save image as
and open it locally
Steps to reproduce:
git clone https://github.com/s0rg/redis-cluster-compose.git
cd redis-cluster-compose
docker compose up -d
then:
decompose -format dot | dot -Tsvg > redis-cluster.svg