Metrics: where and howgraphite-oriented story
• Vsevolod Polyakov• Platform Engineer at Grammarly
GraphiteAll whisper-based systems
Default graphite architecture
what?• RRD-like (gram.ly/gfsx)• so.it.is.my.metric → /so/it/is/my/metric.wsp• Fixed retention (by name\pattern)• Fixed size (actually no)
Retention and size• 1s:1d → 1 036 828 bytes• 10s:10d → 1 036 828 bytes• 1s:365d → 378 432 028 bytes (1 TB ~ 3 000)• 10s:365d → 37 843 228 bytes (1 TB ~ 30 000)
whisper calc
Retention and size• 10s:30d,1m:120d,10m:365d → 4 564 864 bytes• 240 864 metrics in 1 TB• aggregation: average, sum, min, max, and last.• can be assign per metric
How• terraform (https://www.terraform.io/)• docker (https://www.docker.com/)• ansible (https://www.ansible.com/)• rocker (https://github.com/grammarly/rocker)• rocker-compose (
https://github.com/grammarly/rocker-compose)
Default graphite architecture
carbon-cache.py• single-core• many options in config file• default
link
architecturecarbon-cache.py
Start load testing• m4.xlarge instance (4 CPU, 16 GB ram, 256 GB disk EBS gp2)• retentions = 1s:1d• MAX_CACHE_SIZE, MAX_UPDATES_PER_SECOND, MAX_CREATES_PER_MINUTE = inf• defaults• almost 1.5h to get limit :(
carbon-cache.py cache size → 75k req\s
results
• 75 000 req\s max• 60 000 req\s flagman speed• I\O :(
Try to tune!
• WHISPER_SPARSE_CREATE = true (don’t allocate space on creation) non-linear I\O load.
• CACHE_WRITE_STRATEGY = sorted (default)
cache size 1k → 195k req\s
results
• 120 000 req\s flagman speed• cache flush problem :(
Try to tune!
• CACHE_WRITE_STRATEGY = max will give a strong flush preference to frequently updated metrics and will also reduce random file-io.
from 1k to 150k
results
• 90 000 req\s flagman speed• cache flush problem :(
Try to tune!
• CACHE_WRITE_STRATEGY = naive just flush. Better with random I\O.
from 45k to 135k
results
• 120 000 req\s flagman speed• still CPU
sorted
max
naive
• Maybe it’s I\O EBS limitation? → 512 GB disk. • No.
go-carbon• multi-core single daemon• written in golang• not many options to tune :(
link
Start load testing• m4.xlarge instance (4 CPU, 16 GB ram, 256 GB disk EBS gp2)• retentions = 1s:1d• max-size = 0• max-updates-per-second = 0• almost 1h to get limit :(
1k → 130k req\s ~3k/min
results• 120 000 req\s flagman speed• but it’s without sparse. • try to implement
try to tune!remaining := whisper.Size() - whisper.MetadataSize()whisper.file.Seek(int64(remaining-1), 0)whisper.file.Write([]byte{0})chunkSize := 16384zeros := make([]byte, chunkSize)for remaining > chunkSize {
// if _, err = whisper.file.Write(zeros); err != nil {// return nil, err// }remaining -= chunkSize
}if _, err = whisper.file.Write(zeros[:remaining]); err != nil {
return nil, err}
180 000 req\s !
try to tune!
• max update operation = 1500
results
• TLDR 210 000 - 240 000 req\s flagman speed• 31 000 000 cache size!
try to tune!
• max update operation = 0• input-buffer = 400 000
results
• 270 000 req\s flagman speed• 10-20 million req cache size!
try to tune!
• vm.dirty_background_ratio=40• vm.dirty_ratio=60
300 000 req\s
results
• 300 000 req\s flagman speed• 180k+ req\s ±without cache
Re:Lays
Default graphite architecture
arch forward
arch named\regexp
arch hash
arch hash replicafactor: 2
carbon-relay.py
• twisted based• native
Start load testing• c4.xlarge instance (4 CPU, 7.5 GB ram)• ~1 Gb lan• default parameters• hashing• 10 connections
WTF!
carbon-relay-ng• golang-based• web-panel• live-updates• aggregators• spooling
link
<150 000 req\s
carbon-c-relay
• written in C• advanced cluster management
from 100 000 to 1 600 000 req\s
1 400 000 flagman speed. Or not?
So…go-carbon + carbon-c-relay = ♡
BTW. influx, 130k req\s on cluster
influx
openTSDBsingle instance + hbase cluster = upto 150k
req\s
ALSO• zipper:
• https://github.com/grobian/carbonserver• https://github.com/grobian/carbonwriter• https://github.com/dgryski/carbonzipper• https://github.com/dgryski/carbonapi• https://github.com/dgryski/carbonmem
• https://github.com/jssjr/carbonate
plans• Cyanite, retest• newTS• openTSDB tuninig• zipper tuning
feel free to ask• Vsevolod Polyakov• [email protected]• skype: ctrlok1987• github.com/ctrlok• twitter.com/ctrlok• slack: HangOps• Gitter: dev_ua/devops• skype: DevOps from Ukraine