As mentioned in the last post, there are other improvements for Icinga and IDOUtils.
This time, I want to give you a deeper look onto database performance and the housekeeping stuff.
As you might know, selecting, updating or even deleting a row from a table heavily depends on the row count. If table size grows bigger e.g. in the historical tables from IDOUtils, those queries will be slower and hold back the main process. Current approach of IDOUtils is one forked child of ido2db for one idomod connection – working sequentially on the gotten data.
So even one select taking longer will slow down the data processing and worst case the socket will get blocking and idomod complains about writing to data sink.
But how to resolve those issues?
First of all there were several approaches originally found in mysql-mods.sql – setting indexes on table columns which are being used within the WHERE clause. Regarding the fact that ido2db is not just an insert application, but also deletes historical data on demand (table trimming options), selects objects for caching and furthermore updates existing rows (service/hoststatus e.g.) we decided to apply most useful indexes on the table creation statements. It does slow down an insert a bit, but the overall benefit is much bigger than that :-)
Also the upcoming Icinga Web benefits from that – e.g. the logentries tables select performs a lot faster when using the API and a RDBMS.
But that’s not all – indexes are only one approach of improvement. In the last few months, Hendrik, Christoph and myself discussed a lot about the periodic housekeeping. The basic approach was to remove housekeeping function from the main data processing. Simply because historical deletes on large tables will take even longer and prevent new data being written to the database.
There have been discussions about a cronjob and seperated forked processes for housekeeping, but we wanted something within ido2db and simple to use. So Hendrik came up with the idea to create an own thread within each ido2db child which runs completely seperated from the main data processing flow – the so-called threaded housekeeper.
The thread just waits for the appropriate instance getting connected and then performs the periodic housekeeping – independant from the main flow. And it does not interfere with the normal data processing. So to speak it resolves a big performance issue within IDOUtils.
Basically, this is the way it performs:
- sleep a while after creation and intialization
- idle wait for database connection and connected instance from main process
- perform periodic maintenance not interferring with main process
- will be terminated when ido2db shuts down
Best thing so far – it has been implemented and tested and improved quite a while. Mostly done in our own git branches, but the final solution is within current git master and will be one of the outstanding new features for Icinga IDOUtils in the upcoming Icinga 1.0.1 release.
Stay tuned for more updates!
… and prepare for Icinga 1.0.1! =)