We run a web application serving up web APIs for an increasing number of clients. To start, the clients were generally home, office, or other wireless networks submitting chunked http uploads to our API. We've now branched out into handling more mobile clients. The files ranging from a few k to several gigs, broken down into smaller chunks and reassembled on our API.
Our current load balancing is performed at two layers, first we use round robin DNS to advertise multiple A records for our api.company.com address. At each IP, we host a Linux LVS: http://www.linuxvirtualserver.org/, load-balancer that looks at the source IP address of a request to determine which API server to hand the connection to. This LVS boxes are configured with heartbeatd to take-over external VIPs and internal gateway IPs from one another.
Lately, we've seen two new error conditions.
The first error is where clients are oscillating or migrating from one LVS to another, mid-upload. This in turn causes our load balancers to lose track of the persistent connection and send the traffic to a new API server, thereby breaking the chunked upload across two or more servers. Our intent was for the Round Robin DNS TTL value for our api.company.com (which we've set at 1 hour) to be honored by the downstream caching nameservers, OS caching layers, and client application layers. This error occurs for approximately 15% of our uploads.
The second error we've seen much less commonly. A client will initiate traffic to an LVS box and be routed to realserver A behind it. Thereafter, the client will come in via a new source IP address, which the LVS box does not recognize, thereby routing ongoing traffic to realserver B also behind that LVS.
Given our architecture as described in part above, I'd like to know what are people's experiences with a better approach that will allow us to handle each of the error cases above more gracefully?
Edit 5/3/2010:
This looks like what we need. Weighted GSLB hashing on the source IP address.
Best Answer
The canonical solution to this is to not rely on end user IP address, but instead use a Layer 7 (HTTP/HTTPS) load balancer with "Sticky Sessions" via a cookie.
Sticky sessions means the load balancer will always direct a given client to the same backend server. Via cookie means the load balancer (which is itself a fully capable HTTP device) inserts a cookie (which the load balancer creates and manages automagically) to remember which backend server a given HTTP connection should use.
The main downside to sticky sessions is that beckend server load can become somewhat un-even. The load balancer can only distribute load fairly when new connections are made, but given that existing connections may be long-lived in your scenario, then in some time periods load will not be distributed entirely fairly.
Just about every Layer 7 load balancer should be able to do this. On Unix/Linux, some common examples are nginx, HAProxy, Apsis Pound, Apache 2.2 with mod_proxy, and many more. On Windows 2008+ there is Microsoft Application Request Routing. As appliances, Coyote Point, loadbalancer.org, Kemp and Barracuda are common in the low-end space; and F5, Citrix NetScaler and others in high-end.
Willy Tarreau, the author of HAProxy, has a nice overview of load balancing techniques here.
About the DNS Round Robin:
It will not be. And DNS Round Robin isn't a good fit for load balancing. And if nothing else convinces you, keep in mind that modern clients may prefer one host over all others due to longest prefix match pinning, so if the mobile client changes IP address, it may choose to switch to another RR host.
Basically, it's okay to use DNS round robin as a coarse-grained load distribution, by pointing 2 or more RR records to highly available IP addresses, handled by real load balancers in active/passive or active/active HA. And if that's what you're doing, then you might as well serve those DNS RR records with long Time To Live values, since the associated IP addresses are highly available already.