26 October 2015

Cluster strategies in Oracle Hyperion Financial Management

While poking around HFM for other reasons, I’ve found a bunch of interesting constants...
  • STR_HITREG_SERVER_STRATEGY = "ServerStrategy";
From other code and messages, it looks like ServerStrategy is supposed to be a property of the HFM database node in EPM Registry. It’s retrieved on startup, and will always default to 0 (”round robin”) because the property is not actually there unless you manually create it. It lseems to govern the load-distribution strategy employed by clients (i.e. the HFM web application), i.e. where to send each new login if the user has not logged on before and/or the documented StickyServer option is not enabled.

Don't raise your expectations anyway: it looks like the ”load based” strategy (2) has not actually been implemented yet, and if you choose it you’ll actually get the Round-Robin one; whereas ”Random” (1) is really random.

From what I can gather, Round-Robin will follow the list of servers specified in property serverList of the cluster node in EPM Registry, going through the list in the specified order. It’s also entirely in memory and never saved to disk or DB, from which we can deduce that:
  • each webapp or client will do its own thing, with no coordination between them;
  • each webapp or client will forget everything on restart;
  • each webapp will default to Round-Robin strategy and send logins to each appserver in the order specified by serverList
The main advice at this point is hence to keep your beefier appservers at the beginning of serverList, since they're more likely to receive traffic. Alternatively, add ServerStrategy and trust the random algorithm to distribute logins... randomly ;)

I suspect this stuff is a prelude of things to come, because it’s used in sections of code that also deal with SSL initialisation for communication with the appserver, i.e. something that is not quite baked yet but must have been added recently. It’s also pure Java, so can’t be older than Does that mean we’ll finally get real load-balanced clustering in HFM (more than 10 years after people asked for it)…?

No comments: