Jon Probably some details I don't quite understand yet so responses here are to get there...
The concept for scheduling is interesting. Does this basically work around the fact that we have an unfair scheduler so this has queue implementations which pretend data is not available when it knows that there is higher priority data available elsewhere thus returning more threads to the pool faster to increase the likelihood that queues with higher priority data will get served more often? The notion of prioritization implies there is a sorting function happening somewhere. NiFi now does sorting on insertion to every queue. At what points are you suggesting sorting can be done/reduced to? Administration: The existing model does require each prioritizer to be set for each queue. Yours does as well - to opt into this you'd have to select the BucketPrioritizer right? It seems like you're saying the priority ruleset would be provided at the flow controller level and be enforced by all connections which leverage this prioritizer. For large multi-tenant nifi flows having a global ruleset might be too limiting but maybe we just dont worry about that yet. How does this idea work with the fact that queues as the reach a given threshold have their data swapped out to disk and as data gets worked off the flowfiles get swapped back into memory? Thanks Joe On Thu, Oct 17, 2019 at 7:28 AM Kessler, Jon <[email protected]> wrote: > I want to start a discussion about a new prioritization mechanism that > addresses some of the issues that I believe exist in the current solution. > These issues are: > > - Scheduling: No consideration is given to data priority when determining > which component is given the next available thread with which to work > - Constant sorting: Because all flowfiles in a given connection share the > same PriorityQueue they must be sorted every time they move. While this > sort is efficient it can add up as queues grow deep. > - Administration: There is a costly human element to managing the value > used as a priority ranking as priorities change. You must also ensure every > connection in the appropriate flow has the proper prioritizer assigned to > it to make use of the property. > > We have developed a prototype of a new FlowFileQueue implementation that > addresses these issues. Use of this implementation is controlled via > nifi.properties so you can opt-in or out system-wide without doing a lot of > configuration. Its design goals are: > > - Instead of using the value of a FlowFile attribute as a ranking, > maintain a set of expression language rules to define your priorities. The > highest ranked rule that a given FlowFile satisfies will be that FlowFile's > priority > - Because we have a finite set of priority rules we can utilize a bucket > sort in our connections. One bucket per priority rule. The bucket/rule with > which a FlowFile is associated with will be maintained so that as it moves > through the system we do not have to re-evaluate that Flowfile against our > ruleset unless we have reason to do so. > - Control where in your flow FlowFiles are evaluated against the ruleset > with a new Prioritizer implementation: BucketPrioritizer. > - When this queue implementation is polled it will be able to check > state to see if any data of a higher priority than what it currently > contains recently (within 5s) moved elsewhere in the system. If higher > priority data has recently moved elsewhere, the connection will only > provide a FlowFile X% of the time where X is defined along with the rule. > This allows higher priority data to have more frequent access to threads > without thread-starving lower priority data. > - Rules will be managed via a menu option for the flow and changes to > them take effect instantly. This allows you to change your priorities > without stopping/editing/restarting various components on the graph. > > I intend to contribute this solution but first want to solicit input and > opinions. > > - Jon Kessler > >
