The Orion sidecar tale of woe

[c0c0n3]

Adding an Istio sidecar to Orion's pod makes the service unreachable. Also it makes it real hard to do any service initialisation through K8s init containers. Sob, sob.

After so much grief trying to figure out why Envoy doesn't play well with Orion, I've decided to park this issue here for now and deal with it some other time to avoid dev paralysis. What's happening is that, with a sidecar, any HTTP call to Orion from another host, be that inside or outside the mesh, fails with a 503. Requests actually hit the sidecar but then something goes wrong when forwarding to Orion which never actually gets the request and ultimately Envoy returns a 503 to the client. Notice that local HTTP calls actually work, i.e. if you log in to the Orion pod and e.g.

$ curl -v http://localhost:1026/v2

you'll get the expected response from Orion.

One thing to consider is that deploying Orion without a sidecar may actually be an option. In fact:

• The adapter can still intercept requests to Orion, extract the IDS header, do token validation, and ditch or authorise the request as the case may be.
• The Envoy filter won't work since there's no sidecar. So when Orion makes a request to a host external to the mesh, the IDS header can't be added to the request. But if we find a solution to #24, then the adapter will be able to intercept outgoing requests too, so problem solved. And we can ditch the Envoy filter too which is a bit of a hack anyway.
• We can use a simple init container to create DB indexes, etc. With a sidecar in between, we'd need to come up with a real clever init process instead...

[chicco785]

• indexes are for mongo, so that's not much an issue if you don't put a sidecar in front of it but only in front of orion.
• the fact that sidecar is not working, must be due to some configuration issue i suppose. if not all istio has some big issue...

[c0c0n3]

indexes are for mongo, so that's not much an issue if you don't put a sidecar in front of it but only in front of orion.

I tried to put a sidecar in front of MongoDB but couldn't get it right. Orion would keep on crashing trying to connect to the DB. Regarding indexes, I tried to use an init container to run a modified version of the Orchestra script in the Orion chart, but that turned out to be a bit of a mission. So yes, glad to know we can do without :-)

the fact that sidecar is not working, must be due to some configuration issue i suppose. if not all istio has some big issue...

ha :-) yes, definitely. It must be config. I've tried everything I could think of but have had no luck so far.

[c0c0n3]

Sidecar broken wheel

TL;DR. There's likely to be a bug in Istio (1.4 series) that stops sidecars from forwarding HTTP traffic to target pods when naming K8s service/container ports. Workaround: name your service/container port http, then traffic gets forwarded properly.

This is true when using Istio with our config in deployment, I didn't have time to test other scenarios in isolation. Gory details below.

Config

We have a plain K8s service for Orion with a port name of ngsi

...
kind: Service
...
spec:
    type: NodePort
    ports:
    - name: ngsi
      port: 1026
      protocol: TCP
      targetPort: 1026
...

and a simple deployment with a matching container port name

...
kind: Deployment
...
  containers:
  - image: "fiware/orion:2.2.0"
    ports:
    - containerPort: 1026
      name: ngsi
    ...

We also have Istio config to inject a sidecar and have edited the K8s istio-ingressgateway service to make port 31026 accessible from outside the cluster and remap it to port 1026 inside the cluster:

ports:
 ...
   - name: orion
     nodePort: 31026
     port: 1026
     protocol: TCP
     targetPort: 1026

Sidecar init

The Orion sidecar container gets prepped by an Istio init container which sets up iptables to route incoming and outgoing traffic through the Envoy sidecar. There's no configured rules affecting our port in any specific way, e.g. drop packet if port 1026. To see that, log onto the node (Minikube in my case)

$ minikube ssh

Grab the sidecar container ID

$ docker ps | grep orion
0fab7c534b44  ...  k8s_istio-proxy_orion-5c7...
7f50ef4d7931  ...  k8s_orion_orion-5c7...
bd25934a62f1  ...  k8s_POD_orion-5c7...
...

Then use it to figure out its PID

$ docker inspect 0fab7c534b44 --format '{{ .State.Pid }}'
12539

and finally make iptables spill the beans

$ sudo nsenter -t 12539 -n iptables -t nat -S 
...

(more about sidecar injection on the Istio blog.)

Another thing to note is that the Envoy proxy gets started with an initial config where you can see a valid definition for our ngsi port:

$ kubectl exec -it -c istio-proxy orion-5c7d68db9b-4tqvz -- bash
$ ps -F f -A
...
... /usr/local/bin/envoy -c /etc/istio/proxy/envoy-rev0.json ...
...
$ grep ngsi /etc/istio/proxy/envoy-rev0.json
 "metadata": { ...
    "POD_PORTS": [{"name": "ngsi", "containerPort": 1026, "protocol": "TCP"}],
...

Observed behaviour

If you try hitting Orion's /v2 API entry point

curl http://$(minikube ip):31026/v2  \
  -H header:$HEADER_VALUE    # set up HEADER_VALUE as explained in README

you'll get a 503 Service Unavailable back. Here's what happens under the bonnet:

1. Gateway routes request to Mixer.
2. Mixer asks adapter to validate.
3. Adapter okays it.
4. Gateway reroutes request to Orion's sidecar.
5. Sidecar buffers request and establishes TCP connection with Orion---TCP 3-way handshake succeeds but sidecar won't send any data yet.
6. Sidecar closes TCP connection with gateway.
7. Sidecar closes TCP connection with Orion.
8. Gateway returns 503 to curl.

You can trace all that with tcpdump. For example here's how to do that in Orion's sidecar container:

$ kubectl exec -it -c istio-proxy orion-5c7d68db9b-4tqvz -- bash
$ sudo tcpdump -i any -s 1024 -A port 1026

Istio Envoy containers come with tcpdump but you'll have to install it elsewhere.

Considerations

Why does the sidecar slam the door closed, i.e. steps 6 and 7 above? Hard to tell---but how rude :-) One thing is for sure though: this behaviour depends on K8s service/container port names. In fact, renaming our port from ngsi to http solved the problem, i.e. Envoy sends the request on to Orion. (You can try this with the httpbin service too: rename the port to silly and all hell breaks loose.) Also, it doesn't look like our adapter can possibly have anything to do with it, message routing from the gateway to the sidecar works fine, iptables config on the node shouldn't cause any trouble either, so? Well, if I had to point the finger, I'd say there's something wonderfully weird going on in the (Istio modified version of) Envoy.

[c0c0n3]

I think I've got to the bottom of this

https://archive.istio.io/v1.4/docs/ops/configuration/traffic-management/protocol-selection/

even if not explicitly stated, I don't think you can actually use a port name other than those listed for manual protocol selection. If you do, as we saw earlier, the sidecar shuts down the connection. I see a few problems with this:

1. The sidecar interprets K8s port names as protocol identifiers. This effectively changes the semantics of K8s service descriptors.
2. If the sidecar can't figure out the protocol from the port name, it closes the connection instead of falling back to plain TCP relaying---as it should according to the docs, well at least that's my interpretation.
3. Nothing gets logged, so you're left looking for a needle in a haystack.

I'd argue that from a design standpoint (1) isn't the most fortunate choice, but will let philosophers judge that on its own merits. As for (2), it smells like a bug to me and (3) is something the Istio guys definitely need to improve, IMHO.

[c0c0n3]

Start-up race conditions

A slow-starting sidecar could stop Orion from completing successfully its start-up procedure.

When injecting a sidecar, Istio also adds an init container to set up iptables rules to redirect inbound and outbound traffic to the sidecar. After executing the init container, K8s starts the sidecar and Orion concurrently. As part of its start-up procedure, Orion tries to establish a connection to MongoDB and, since iptables rules are already in place at this stage, the kernel redirects those TCP packets to the sidecar. If the sidecar isn't ready yet to process messages, packets get dropped. When this happens, I've observed the following outcomes:

1. Orion deadlocks. After running out of connection retries, Orion seems to just sit there doing nothing, the process is still alive but becomes unresponsive, so the container won't terminate.
2. Orion eventually recovers. All is well.

We should make it clear to mesh admins that to prevent (1) from happening it's probably a good idea for their K8s configuration to:

1. Delay slightly Orion startup---e.g. sleep 2 && /usr/bin/contextBroker ...
2. Use readiness/liveness probes---e.g. liveness on /v2.

Also, Orion logs say DB connections get retried 100 times with a 1000 microsecond delay in between, e.g.

Database Startup Error 
(cannot connect to mongo - doing 100 retries with a 1000 microsecond interval)

That would mean only 0.1 milliseconds between the first and last attempt which seems to be a bit too little in a mesh environment---e.g. MongoDB could be much slower than that at start up. Perhaps there's a way to configure Orion so that the time gap between retries is a bit wider? Not sure how much the CLI option -dbTimeout could help here...

[c0c0n3]

closed by #36