couchbase/cbmonitor

proposal to change metric format/querying

Closed this issue · 4 comments

mschoch commented

I see query for one of our metrics going to seriesly as:

http://localhost:3133/cbmonitor/_query?group=1000&ptr=/metric/laptop/beer-sample/127.0.0.1/curr_items_tot&reducer=avg

In my seriesly db this returns data like:

{
"1361547184000": [
null
],
"1361547194000": [
null
],
"1361547204000": [
null
],
"1361547214000": [
null
],
"1361547225000": [
null
],
"1361547227000": [
7303
],
"1361547235000": [
null
],
"1361547237000": [
7303
],
"1361547245000": [
null
],
"1361547248000": [
7303
],
"1361547255000": [
null
],
"1361547259000": [
7303
],
"1361547265000": [
null
],
"1361547269000": [
7303
],
"1361547275000": [
null
],
"1361547280000": [
7303
],
"1361547285000": [
null
],
"1361547291000": [
7303
],
"1361547295000": [
null
],
"1361547301000": [
7303
],
"1361547305000": [
null
],
"1361547312000": [
7303
],
"1361547315000": [
null
],
"1361547323000": [
7303
],
"1361547326000": [
null
],
"1361547335000": [
7303
],
"1361547346000": [
7303
]
}

I believe the large number of nulls is because I have many different kinds of metrics being stored (also from multiple clusters).

What if we stored data more like:

{
"meta": {
"cluster": cluster-name (optional)
"node": node-name (optional)
"bucket": bucket-name (optional)
},
"metric": {
...
}
}

Then in we'd use both ptr and filters in the query:

ptr=/metric/curr_items_tot
f=/meta/cluster
fv=laptop
f=/meta/bucket
vb=beer-sample
f=/meta/node
fv=127.0.0.1

This would also be more extendable for different kinds of metric meta-data in the future.

Also would like to ping @dustin for input on this.

pavel-paulau commented

Current approach is little bit ugly because I was trying to avoid filters in queries.
From API perspective filters are definitively more convenient and flexible but their performance...

mschoch commented

One of the other reasons I like this approach is that you could discover the clusters/nodes/buckets through seriesly:

Clusters:
/cbmonitor/_query?ptr=/meta/cluster&reducer=identity&group=1000000000

Nodes:
/cbmonitor/_query?ptr=/meta/node&reducer=identity&group=1000000000

Buckets:
/cbmonitor/_query?ptr=/meta/bucket&reducer=identity&group=1000000000

We could also have the collectors record metric meta-data differently, and lose the need for the other database.

I know I'm proposing a bold change here, but I think it would greatly simplify using the tool. It would also mean different frontends could all easily sit on top of seriesly to visualize this.

mschoch commented

Hmmn, I'll talk to Dustin about filters to understand their impact better.