loadbalancer for hstreamdb

[mntnclmbr]

Hi,
Do you recommended a load balancer in front of a 3 node cluster? In AWS, I was thinking of setting up an application load balancer in front of the cluster so the incoming connections are automatically distributed.

[alissa-tung]

Hello, as for load balancing, there is no need to set up a load balancer n front of the HStream cluster.

Our server and client implementation had already done load balancing stuff, it will distrust connections over the cluster of servers with its load balancing algorithms and methods. Hence, there is no need for extra load balancer and it can not (since HStream has its internal strategy) to set up one.

[mntnclmbr]

if we don't use a LB, what should the url be for sending data to hstream cluster ?

[alissa-tung]

If you are using any client implementation which needs a URL to establish a client connection, the answer is anyone in the cluster.

When doing resource dependent operations, the client implementation will first do lookup the resource, then send the operation request to the correct server. (FYI, but there is no need for users to know and consider the implementation details). In addition, the client library implementation is suggested to keep a connection pool managed by itself in its internal state so that the switch of URLs is seamless and not transparent to client library users.

[mntnclmbr]

I assume the servers will each maintain a list of all the other nodes in the cluster. What I am wondering is for a cluster on AWS, this would mean the IP it tracks has to be the external IP (for AWS EC2 instances).

So the IPs in the config.yaml in https://github.com/hstreamdb/deployment-tool , for section :
hserver:

• host: 172.24.47.173
• host: 172.24.47.174
• host: 172.24.47.175
  will have to be all external IPS. I was planning to do it from another machine that would have been inside the VPC and would have access to the internal IPs but if the hserver does a redirect to another node, it will fail since the external client won't be able to access the internal IP.

This also means the data between the nodes will always have to go out of the VPC and access the node via the external IP . This would potentially not be a security issue(assuming TLS connection between the nodes), but will incur additional latency.

Am I thinking about it right or am I missing something ?

As an example for my EMQX cluster, I have set them up to only use the internal IP of the VPC so no inter-node data needs to go out of the VPC and is all within the local network. It uses a load balancer in front to distribute the incoming requests to each of the node itself, with no redistribution happening on the actual node.

Also, is there any other way to establish a client connection that does not use the URL ? I was looking at https://github.com/hstreamdb/hstreamdb-py for this.

Sorry if I am asking dumb questions and do appreciate all the help!

[alissa-tung]

• When setting up HStream cluster in the cloud, it is recommend to use the deployment tool on one machine inside the VPC. So the config is expected to be filled with private (inside VPC) IPs. The IPs (or to say, any URL can finally provide host and port) here will also be used in generating the HServer cluster config and HStore cluster config so using IPs inside VPC is also for effective considerations.
• To connect every HServers without exposing all IP is indeed a production scene. We are developing a components HStream Proxy which is like a gateway, can archive the goal to just expose one endpoint. I think it can be used for your purpose. We will open source it later, and please feel free to ask any questions about at anytime!
• If only EMQX integration is used in connecting HStream cluster, it is recommend to setup EMQX and HStream inside one VPC.

[alissa-tung]

Also, is there any other way to establish a client connection that does not use the URL ? I was looking at https://github.com/hstreamdb/hstreamdb-py for this.

The API for connecting is here.
To establish a connection in Python client, we need either

• host + port
• URL, and if you provide this url then the ‘host’ and ‘port’ args will be ignored.

I am not sure what does not using the URL is about. AFAIK you can ignore the url argument (it has default to None) once the host and port is set. Anyway we just need one thing to get the server destination.

cc @4eUeP