[ZBXNEXT-4072] Spread out trend flushing in time Created: 2017 Aug 31 Updated: 2019 Dec 11 Resolved: 2017 Nov 23 |
|
| Status: | Closed |
| Project: | ZABBIX FEATURE REQUESTS |
| Component/s: | Server (S) |
| Affects Version/s: | 4.0.0alpha1 |
| Fix Version/s: | None |
| Type: | Change Request | Priority: | Major |
| Reporter: | Glebs Ivanovskis (Inactive) | Assignee: | Unassigned |
| Resolution: | Commercial support required | Votes: | 2 |
| Labels: | load, performance, trends | ||
| Remaining Estimate: | Not Specified | ||
| Time Spent: | Not Specified | ||
| Original Estimate: | Not Specified | ||
| Issue Links: |
|
||||||||
| Description |
|
Currently trend cache can hold only one entry per item meaning that at the beginning of every hour when item receives the first value for this hour item trend for the previous hour has to be flushed to DB immediately to start accumulating new trend in the cache. Given that most items have update intervals which are less then 5 minutes such approach to trend caching leads to peak loads for history syncers and DB backend in first five minutes of an hour. This slows down overall Zabbix data flow and introduces latency in problem detection and escalation. To counter this effect trend cache should be redesigned to be able to store more than one trend per item and flush trends of past hours without a rush. |
| Comments |
| Comment by Oleksii Zagorskyi [ 2017 Aug 31 ] |
|
It's especially hard for very first trends calculation after daemon restart. Because zabbix server needs to check possibly existing records for past hour in trends and update them. |
| Comment by Vladislavs Sokurenko [ 2017 Sep 01 ] |
|
What do you mean by peak load ? Are hystory syncers actually loaded and doing something useful or just waiting on one another ? if waiting, then on which table they spend time waiting ? You mentioned something that system is not usable, is it because of load or because some tables are locked ? |
| Comment by Glebs Ivanovskis (Inactive) [ 2017 Sep 01 ] |
|
I mean history syncer busyness graphs like in https://www.zabbix.com/forum/showthread.php?t=54370 Waiting for DB. Inserts into history* are OK, inserts into trends* are slower. |
| Comment by Vladislavs Sokurenko [ 2017 Sep 01 ] |
|
Note sure if this could be related somehow, maybe not: but if problem is reproducible then should be easy to test |
| Comment by Glebs Ivanovskis (Inactive) [ 2017 Sep 01 ] |
|
Definitely not. What if there are many items with one hour update interval? |
| Comment by bunkzilla [ 2019 Dec 11 ] |
|
this can pretty much cause a pagerstorm from triggers I have configured for nodata because of the slowdown encountered at the 00 minute when the trends are being managed.
|