fabio is a fast, modern, zero-conf load balancing HTTP(S) router for deploying microservices managed by consul.
Both hardware and software routers like Citrix Netscaler, F5 Big IP, haproxy, varnish, nginx or apache require some form of configuration - the routing table - to route incoming traffic to services which can handle them. The routing table has to be kept in sync with the actual deployed set of services and instances during each deployment and each outage on every environment. This makes the routing table a crucial part of the configuration without the application cannot function.
Managing the routing table can be automated via API calls or tools like consul-template but that also requires configuration and/or tools. In the case of consul-template the config file template itself has to be kept in sync with the actual setup of the application. Finally, updating the routing table without loss of existing connections can be [challenging](http://engineeringblog.yelp.com/2015/04/true- zero-downtime-haproxy-reloads.html).
fabio solves this problem by making the services themselves responsible for updating the routing table. Services already know which routes they serve since they have handlers who can handle requests for them. Once services push the routes they handle into the service registry (in this case consul) fabio can build the routing table and can re-configure itself on every change automatically without restart and without the loss of existing connections.
The motivation is also outlined in the presentation I've given at the dotGo EU pre-party in Paris on 9 Nov 2015. You can watch it here.
fabio was developed at the eBay Classifieds Group in Amsterdam and routes traffic for marktplaats.nl and kijiji.it. Marktplaats is running all of its traffic through fabio which is several thousand requests per second distributed over several fabio instances. We don't observe any measurable additional latency.
- Single binary in Go. No external dependencies.
- Zero-conf
- Hot-reloading of routing table through backend watchers
- Round robin and random distribution
- Traffic Shaping (send 5% of traffic to new instances)
- Graphite metrics
- Request tracing
- WebUI
- Fast
- v1.0.4: SSL client certificate authentication support (see
proxy.addr
in fabio.properties) - v1.0.5:
X-Forwarded-For
andForwarded
header support - v1.0.5: Websocket support (experimental)
- v1.0.6: Raw websocket support as default
- v1.0.6: Experimental HTTP api
- v1.0.6: Improved UI
- v1.0.6: fabio registers itself in consul
- Installation
- Quickstart
- Deployment
- Configuration (documented fabio.properties file)
- Performance
- Service configuration
- Manual overrides
- Routing
- Traffic shaping
- Websockets
- Debugging
- Request tracing
- Web UI
- Changelog
- License
This is how you use fabio in your setup:
- Register your service in consul
- Register a health check in consul as described here. Make sure the health check is passing since fabio will only watch services which have a passing health check.
- Register one
urlprefix-
tag perhost/path
prefix it serves, e.g.urlprefix-/css
,urlprefix-/static
,urlprefix-mysite.com/
- Start fabio without a config file (assuming a consul agent on
localhost:8500
) Watch the log output how fabio picks up the route to your service. Try starting/stopping your service to see how the routing table changes instantly. - Send all your HTTP traffic to fabio on port
9999
- Done
To start a sample server to test the routing run the demo/server
like this:
./server -addr 127.0.0.1:5000 -name svc-a -prefix /foo
and access the server direct and via fabio
curl 127.0.0.1:5000/foo # direct
curl 127.0.0.1:9999/foo # via fabio
If you want fabio to handle SSL as well set the proxy.addr
along with the
public/private key files in
fabio.properties
and run fabio -cfg fabio.properties
. You might also want to set the
proxy.header.clientip
, proxy.header.tls
and proxy.header.tls.value
options.
Check the Debugging section to see how to test fabio with curl
.
See fabio in action
[](https://www.youtube.com/watch?v=gvxxu0PLevs"fabio demo - Click to Watch!")
The fabio-example
project is now in the demo/server
directory.
To install fabio run (you need Go 1.4 or higher)
go get github.com/eBay/fabio
To start fabio run
./fabio
which will run it with the default configuration which is described
in fabio.properties
. To run it with a config file run it
with
./fabio -cfg fabio.properties
or use the official Docker image and mount your own config file to /etc/fabio/fabio.properties
docker run -d -p 9999:9999 -p 9998:9998 -v $PWD/fabio/fabio.properties:/etc/fabio/fabio.properties magiconair/fabio
If you want to run the Docker image with one or more SSL certificates then
you can store your configuration and certificates in /etc/fabio
and mount
the entire directory, e.g.
$ cat ~/fabio/fabio.properties
proxy.addr=:443;/etc/fabio/ssl/mycert.pem;/etc/fabio/ssl/mykey.pem
docker run -d -p 443:443 -p 9998:9998 -v $PWD/fabio:/etc/fabio magiconair/fabio
The official Docker image contains the root CA certificates from a recent and updated Ubuntu 12.04.5 LTS installation.
The main use-case for fabio is to distribute incoming HTTP(S) requests from the internet to frontend (FE) services which can handle these requests. In this scenario the FE services then use the service discovery feature in consul to find backend (BE) services they need in order to fulfil the request.
That means that fabio is currently not used as an FE-BE or BE-BE router to route traffic among the services themselves since the service discovery of consul already solves that problem. Having said that, there is nothing that inherently prevents fabio from being used that way. It just means that we are not doing it.
In the following setup fabio is configured to listen on the public ip(s) where it can optionally terminate SSL traffic for one or more domains - one ip per domain.
+--> service-a
|
internet -- HTTP/HTTPS --> fabio -- HTTP --+--> service-b
|
+--> service-c
To scale fabio you can deploy it together with the frontend services which provides high-availability and distributes the network bandwidth.
+- HTTP/HTTPS -> fabio -+- HTTP -> service-a (host-a)
| |
internet --+- HTTP/HTTPS -> fabio -+- HTTP -> service-b (host-b)
| |
+- HTTP/HTTPS -> fabio -+- HTTP -> service-c (host-c)
In the following setup fabio is configured receive all incoming traffic from an existing gateway which also terminates SSL for one or more domains. fabio supports SSL Client Certificate Authentication to support the Amazon API Gateway
+--> service-a
|
internet -- HTTP/HTTPS --> LB -- HTTP --> fabio -- HTTP --+--> service-b
|
+--> service-c
Again, to scale fabio you can deploy it together with the frontend services which provides high-availability and distributes the network bandwidth.
+- HTTP -> fabio -+-> service-a (host-a)
| |
internet -- HTTP/HTTPS --> LB -+- HTTP -> fabio -+-> service-b (host-b)
| |
+- HTTP -> fabio -+-> service-c (host-c)
fabio has been tested to deliver up to 15.000 req/sec on a single 16 core host with moderate memory requirements (~ 60 MB).
To achieve the performance fabio sets the following defaults which can be overwritten with the environment variables:
GOMAXPROCS
is set toruntime.NumCPU()
since this is not the default for Go 1.4 and beforeGOGC=800
is set to reduce the pressure on the garbage collector
When fabio is compiled with Go 1.5 and run with default settings it can be up
to 40% slower than the same version compiled with Go 1.4. The GOGC=100
default puts more pressure on the Go 1.5 GC which makes the fabio spend 10% of
the time in the GC. With GOGC=800
this drops back to 1-2%. Higher values
don't provide higher gains.
As usual, don't rely on these numbers and perform your own benchmarks. You can
check the time fabio spends in the GC with GODEBUG=gctrace=1
.
Each service can register one or more URL prefixes for which it serves
traffic. A URL prefix is a host/path
combination without a scheme since SSL
has already been terminated and all traffic is expected to be HTTP. To
register a URL prefix add a tag urlprefix-host/path
to the service
definition.
By default, traffic is distributed evenly across all service instances which register a URL prefix but you can set the amount of traffic a set of instances will receive ("Canary testing"). See [Traffic Shaping](#Traffic Shaping) below.
A background process watches for service definition and health status changes in consul. When a change is detected a new routing table is constructed using the commands described in [Config Commands](#Config Commands).
Since an automatically generated routing table can only be changed with a service deployment additional routing commands can be stored manually in the consul KV store which get appended to the automatically generated routing table. This allows fine-tuning and fixing of problems without a deployment.
The [Traffic Shaping](#Traffic Shaping) commands are also stored in the KV store.
The routing table is configured with the following commands:
route add <svc> <src> <dst> weight <w> tags "<t1>,<t2>,..."
- Add route for service svc from src to dst and assign weight and tags
route add <svc> <src> <dst> weight <w>
- Add route for service svc from src to dst and assign weight
route add <svc> <src> <dst> tags "<t1>,<t2>,..."
- Add route for service svc from src to dst and assign tags
route add <svc> <src> <dst>
- Add route for service svc from src to dst
route del <svc> <src> <dst>
- Remove route matching svc, src and dst
route del <svc> <src>
- Remove all routes of services matching svc and src
route del <svc>
- Remove all routes of service matching svc
route weight <svc> <src> weight <w> tags "<t1>,<t2>,..."
- Route w% of traffic to all services matching svc, src and tags
route weight <src> weight <w> tags "<t1>,<t2>,..."
- Route w% of traffic to all services matching src and tags
route weight <svc> <src> weight <w>
- Route w% of traffic to all services matching svc and src
route weight service host/path weight w tags "tag1,tag2"
- Route w% of traffic to all services matching service, host/path and tags
w is a float > 0 describing a percentage, e.g. 0.5 == 50%
w <= 0: means no fixed weighting. Traffic is evenly distributed
w > 0: route will receive n% of traffic. If sum(w) > 1 then w is normalized.
sum(w) >= 1: only matching services will receive traffic
Note that the total sum of traffic sent to all matching routes is w%.
The order of commands matters but routes are always ordered from most to least specific by prefix length.
The routing table contains first all routes with a host sorted by prefix length in descending order and then all routes without a host again sorted by prefix length in descending order.
For each incoming request the routing table is searched top to bottom for a
matching route. A route matches if either host/path
or - if there was no
match - just /path
matches.
The matching route determines the target URL depending on the configured
strategy. rnd
and rr
are available with rnd
being the default.
The auto-generated routing table is
route add service-a www.mp.dev/accounts/ http://host-a:11050/ tags "a,b"
route add service-a www.kjca.dev/accounts/ http://host-a:11050/ tags "a,b"
route add service-a www.dba.dev/accounts/ http://host-a:11050/ tags "a,b"
route add service-b www.mp.dev/auth/ http://host-b:11080/ tags "a,b"
route add service-b www.kjca.dev/auth/ http://host-b:11080/ tags "a,b"
route add service-b www.dba.dev/auth/ http://host-b:11080/ tags "a,b"
The manual configuration under /fabio/config
is
route del service-b www.dba.dev/auth/
route add service-c www.somedomain.com/ http://host-z:12345/
The complete routing table then is
route add service-a www.mp.dev/accounts/ http://host-a:11050/ tags "a,b"
route add service-a www.kjca.dev/accounts/ http://host-a:11050/ tags "a,b"
route add service-a www.dba.dev/accounts/ http://host-a:11050/ tags "a,b"
route add service-b www.mp.dev/auth/ http://host-b:11080/ tags "a,b"
route add service-b www.kjca.dev/auth/ http://host-b:11080/ tags "a,b"
route add service-c www.somedomain.com/ http://host-z:12345/ tags "a,b"
fabio allows to control the amount of traffic a set of service instances will
receive. You can use this feature to direct a fixed percentage of traffic to a
newer version of an existing service for testing ("Canary testing"). See
Manual Overrides for a complete description of the route weight
command.
The following command will allocate 5% of traffic to www.kjca.dev/auth/
to
all instances of service-b
which match tags version-15
and dc-fra
. This
is independent of the number of actual instances running. The remaining 95%
of the traffic will be distributed evenly across the remaining instances
publishing the same prefix.
route weight service-b www.kjca.dev/auth/ weight 0.05 tags "version-15,dc-fra"
Websocket support works but is considered experimental since I don't have an in-house use case for it at the moment. I would like to hear from users whether it works in their environments beyond my simple test case before I declare it stable. It has been implemented with the websocket library from golang.org/x/net/websocket
You can test the websocket support with the demo/wsclient
and demo/server
which
implements a simple echo server.
./server -addr 127.0.0.1:5000 -name ws-a -prefix /echo -proto ws
./wsclient -url ws://127.0.0.1:9999/echo
You can also run multiple web socket servers on different ports but the same endpoint.
fabio detects on whether to forward the request as HTTP or WS based on the
value of the Upgrade
header. If the value is websocket
it will attempt a
websocket connection to the target. Otherwise, it will fall back to HTTP.
One limitation of the current implementation is that the accepted set of protocols has to be symmetric across all services handling it. Only the following combinations will work reliably:
svc-a and svc-b register /foo and accept only HTTP traffic there
svc-a and svc-b register /foo and accept only WS traffic there
svc-a and svc-b register /foo and accept both HTTP and WS traffic there
The following setup (or variations thereof) will not work reliably:
svc-a registers /foo and accept only WS traffic there
svc-b registers /foo and accept only HTTP traffic there
This is not a limitation of the routing itself but because the current configuration does not provide fabio with enough information to make the routing decision since the services do not advertise the protocols they handle on a given endpoint.
This does not look like a big restriction but is also not difficult to extend in a later version assuming there are use cases which require this behavior. For now the services have to be symmetric in the protocols they accept.
To send a request from the command line via the fabio using curl
you should send it as follows:
curl -v -H 'Host: foo.com' 'http://localhost:9999/path'
The -x
or --proxy
options will most likely not work as you expect as they
send the full URL instead of just the request URI which usually does not match
any route but the default one - if configured.
To trace how a request is routed you can add a Trace
header with an non-
empty value which is truncated at 16 characters to keep the log output short.
$ curl -v -H 'Trace: abc' -H 'Host: foo.com' 'http://localhost:9999/bar/baz'
2015/09/28 21:56:26 [TRACE] abc Tracing foo.com/bar/baz
2015/09/28 21:56:26 [TRACE] abc No match foo.com/bang
2015/09/28 21:56:26 [TRACE] abc Match foo.com/
2015/09/28 22:01:34 [TRACE] abc Routing to http://1.2.3.4:8080/
fabio contains a simple web ui to examine the routing table and manage the
manual overrides. By default it is accessible on http://localhost:9998/
MIT licensed