Multiple active gateway enhancement proposal

[anfredette]

Submariner currently only allows a single gateway to be active at
any one time in a given cluster. This enhancement proposes that
there be an option to allow more than one gateway to be active in
a given cluster to enhance both performance and resiliency. This
feature is sometimes referred to as “active-active gateways”;
however, the goal is to allow two or more active gateways to be used.

Signed-off-by: Andre Fredette [email protected]

[submariner-bot]

🤖 Created branch: z_pr69/anfredette/multi-gateway

[anfredette]

Thanks for your feedback @astoycos!

[nyechiel]

I have a general comment about this proposal: based on user's feedback so far, it's not clear if the gateway model (where we assume that dedicated nodes or certain worker nodes have public IPs available) is really the future of Submariner. This model is not very cloud-native, cumbersome to setup, and is not compatible/realistic with many of the managed Kubernetes offerings which are becoming more popular. For this reason, we started some early exploration work around using a LoadBalancer type service instaed (see also https://submariner.io/getting-started/architecture/gateway-engine/). I am wondering if we shouldn't complete the investigation of the LoadBalancer mode first, before further investing into HA of the gateway model.

[sridhargaddam]

I have a general comment about this proposal: based on user's feedback so far, it's not clear if the gateway model (where we assume that dedicated nodes or certain worker nodes have public IPs available) is really the future of Submariner. This model is not very cloud-native, cumbersome to setup, and is not compatible/realistic with many of the managed Kubernetes offerings which are becoming more popular. For this reason, we started some early exploration work around using a LoadBalancer type service instaed (see also https://submariner.io/getting-started/architecture/gateway-engine/). I am wondering if we shouldn't complete the investigation of the LoadBalancer mode first, before further investing into HA of the gateway model.

+1
This is inline with what I mentioned earlier on couple of occasions that LB mode is going to be the future.

[mkolesnik]

+1 to @nyechiel's assessment.

Additionally, LB mode would have many of the problems of multiple gateways (since we'd probably want to run more than one GW pod).
I'm suggesting we focus on LB mode first, and that would also pave the way to an easier implementation of multiple dedicated gateways, should such a feature be desired.

[anfredette]

Thanks everyone for the review and comments/suggestions. We've updated the proposal to reflect our current thoughts on the design, and the POC that has been implemented. We're currently doing some performance testing and will update the pr when we get some results.

[Billy99]

Added updates based on the Tech Exchange. Also added some PoC details.

[sridhargaddam]

I see that the open-items/limitations discussed during the tech-exchange session are added as TODO/Notes in the PR. LGTM, thank you.

[mkolesnik]

I see concerns are listed as TODO, so we can merge as "Deferred", with the intention that should this be picked up in the future the entire EP will have to be re-reviewed.

[anfredette]

@sridhargaddam, I think this is good to merge now, isn't it? Thanks, Andre

[sridhargaddam]

@sridhargaddam, I think this is good to merge now, isn't it? Thanks, Andre

Indeed, thank you @anfredette

[submariner-bot]

🤖 Closed branches: [z_pr69/anfredette/multi-gateway]

[anfredette]

Thanks, @sridhargaddam!