Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Get Instrumented: How Prometheus Can Unify Your...
Search
Hynek Schlawack
May 31, 2016
Programming
4
11k
Get Instrumented: How Prometheus Can Unify Your Metrics
Hynek Schlawack
May 31, 2016
Tweet
Share
More Decks by Hynek Schlawack
See All by Hynek Schlawack
Subclassing, Composition, Python, and You
hynek
3
310
Classy Abstractions @ Python Web Conf
hynek
0
190
On the Meaning of Version Numbers
hynek
0
320
Maintaining a Python Project When It’s Not Your Job
hynek
1
2.4k
How to Write Deployment-friendly Applications
hynek
0
2.5k
Solid Snakes or: How to Take 5 Weeks of Vacation
hynek
2
5.8k
Beyond grep – PyCon JP
hynek
1
3.4k
Beyond grep – EuroPython Edition
hynek
1
10k
Beyond grep: Practical Logging and Metrics
hynek
3
1.2k
Other Decks in Programming
See All in Programming
Laravel × Clean Architecture
bumptakayuki
PRO
0
130
エンジニア向けCursor勉強会 @ SmartHR
yukisnow1823
3
12k
ComposeでWebアプリを作る技術
tbsten
0
130
Bedrock×MCPで社内ブログ執筆文化を育てたい!
har1101
7
1.4k
設計の本質:コード、システム、そして組織へ / The Essence of Design: To Code, Systems, and Organizations
nrslib
10
3.7k
RubyKaigi Dev Meeting 2025
tenderlove
1
1.3k
20250429 - CNTUG Meetup #67 / DevOps Taiwan Meetup #69 - Deep Dive into Tetragon: Building Runtime Security and Observability with eBPF
tico88612
0
160
KawaiiLT 登壇資料 キャリアとモチベーション
hiiragi
0
160
Thank you <💅>, What's the Next?
ahoxa
1
590
eBPF超入門「o11yに使える」とは (20250424_eBPF_o11y)
thousanda
1
110
Cline with Amazon Bedrockで爆速開発体験ハンズオン/ 株式会社ブリューアス登壇資料
mhan
0
110
カオスに立ち向かう小規模チームの装備の選択〜フルスタックTSという装備の強み _ 弱み〜/Choosing equipment for a small team facing chaos ~ Strengths and weaknesses of full-stack TS~
bitkey
1
130
Featured
See All Featured
XXLCSS - How to scale CSS and keep your sanity
sugarenia
248
1.3M
The Language of Interfaces
destraynor
157
25k
Why You Should Never Use an ORM
jnunemaker
PRO
56
9.3k
Git: the NoSQL Database
bkeepers
PRO
430
65k
Bash Introduction
62gerente
611
210k
YesSQL, Process and Tooling at Scale
rocio
172
14k
How to Think Like a Performance Engineer
csswizardry
23
1.6k
Exploring the Power of Turbo Streams & Action Cable | RailsConf2023
kevinliebholz
32
5.5k
It's Worth the Effort
3n
184
28k
How GitHub (no longer) Works
holman
314
140k
Java REST API Framework Comparison - PWX 2021
mraible
31
8.5k
Speed Design
sergeychernyshev
29
920
Transcript
Hynek Schlawack Get Instrumented How Prometheus Can Unify Your Metrics
Goals
Goals
Goals
Goals
Goals
Service Level
Service Level Indicator
Service Level Indicator Objective
Service Level Indicator Objective (Agreement)
Metrics
Metrics avg latency 0.3 0.5 0.8 1.1 2.6
Metrics 12:00 12:01 12:02 12:03 12:04 avg latency 0.3 0.5
0.8 1.1 2.6
Metrics 12:00 12:01 12:02 12:03 12:04 avg latency 0.3 0.5
0.8 1.1 2.6 server load 0.3 1.0 2.3 3.5 5.2
None
Instrument
Instrument
Instrument
Instrument
Instrument
None
None
Metric Types
Metric Types ❖counter
Metric Types ❖counter ❖gauge
Metric Types ❖counter ❖gauge ❖summary
Metric Types ❖counter ❖gauge ❖summary ❖histogram
Metric Types ❖counter ❖gauge ❖summary ❖histogram ❖ buckets (1s, 0.5s,
0.25, …)
Averages
❖ avg(request time) ≠ avg(UX) Averages
❖ avg(request time) ≠ avg(UX) ❖ avg({1, 1, 1, 1,
10}) = 2.8 Averages
❖ avg(request time) ≠ avg(UX) ❖ avg({1, 1, 1, 1,
10}) = 2.8 Averages
❖ avg(request time) ≠ avg(UX) ❖ avg({1, 1, 1, 1,
10}) = 2.8 Averages
❖ avg(request time) ≠ avg(UX) ❖ avg({1, 1, 1, 1,
10}) = 2.8 ❖ median({1, 1, 1, 1, 10}) = 1 Averages
❖ avg(request time) ≠ avg(UX) ❖ avg({1, 1, 1, 1,
10}) = 2.8 ❖ median({1, 1, 1, 1, 10}) = 1 Averages
❖ avg(request time) ≠ avg(UX) ❖ avg({1, 1, 1, 1,
10}) = 2.8 ❖ median({1, 1, 1, 1, 10}) = 1 ❖ median({1, 1, 100_000}) = 1 Averages
Percentiles
Percentiles nth percentile P of a data set = P
≥ n% of values
None
50th percentile = 1 ms
50th percentile = 1 ms 50% of requests done by
1 ms
Percentiles
Percentiles P {1, 1, 100_000} 50th 1
Percentiles P {1, 1, 100_000} 50th 1 95th 90_000
None
None
None
Naming
Naming backend1_app_http_reqs_msgs_post backend1_app_http_reqs_msgs_get …
Naming backend1_app_http_reqs_msgs_post backend1_app_http_reqs_msgs_get … app_http_reqs_total
Naming backend1_app_http_reqs_msgs_post backend1_app_http_reqs_msgs_get … app_http_reqs_total
Naming backend1_app_http_reqs_msgs_post backend1_app_http_reqs_msgs_get … app_http_reqs_total
Naming backend1_app_http_reqs_msgs_post backend1_app_http_reqs_msgs_get … app_http_reqs_total{meth="POST", path="/msgs", backend="1"} app_http_reqs_total{meth="GET", path="/msgs", backend="1"}
… app_http_reqs_total
None
None
1. resolution = scraping interval
1. resolution = scraping interval 2. missing scrapes = less
resolution
Pull: Problems ❖ short lived jobs
None
Pull: Problems ❖ short lived jobs ❖ target discovery
Configuration scrape_configs: - job_name: 'prometheus' target_groups: - targets: - 'localhost:9090'
Configuration scrape_configs: - job_name: 'prometheus' target_groups: - targets: - 'localhost:9090'
Configuration scrape_configs: - job_name: 'prometheus' target_groups: - targets: - 'localhost:9090'
Configuration scrape_configs: - job_name: 'prometheus' target_groups: - targets: - 'localhost:9090'
{instance="localhost:9090",job="prometheus"}
None
Pull: Problems ❖ target discovery ❖ short lived jobs ❖
Heroku/NATed systems
Pull: Advantages
Pull: Advantages ❖ multiple Prometheis easy
Pull: Advantages ❖ multiple Prometheis easy ❖ outage detection
Pull: Advantages ❖ multiple Prometheis easy ❖ outage detection ❖
predictable, no self-DoS
Pull: Advantages ❖ multiple Prometheis easy ❖ outage detection ❖
predictable, no self-DoS ❖ easy to instrument 3rd parties
Metrics Format # HELP req_seconds Time spent \ processing a
request in seconds. # TYPE req_seconds histogram req_seconds_count 390.0 req_seconds_sum 177.0319407
Metrics Format # HELP req_seconds Time spent \ processing a
request in seconds. # TYPE req_seconds histogram req_seconds_count 390.0 req_seconds_sum 177.0319407
Metrics Format # HELP req_seconds Time spent \ processing a
request in seconds. # TYPE req_seconds histogram req_seconds_count 390.0 req_seconds_sum 177.0319407
Metrics Format # HELP req_seconds Time spent \ processing a
request in seconds. # TYPE req_seconds histogram req_seconds_count 390.0 req_seconds_sum 177.0319407
Metrics Format # HELP req_seconds Time spent \ processing a
request in seconds. # TYPE req_seconds histogram req_seconds_count 390.0 req_seconds_sum 177.0319407
Percentiles req_seconds_bucket{le="0.05"} 0.0 req_seconds_bucket{le="0.25"} 1.0 req_seconds_bucket{le="0.5"} 273.0 req_seconds_bucket{le="0.75"} 369.0 req_seconds_bucket{le="1.0"}
388.0 req_seconds_bucket{le="2.0"} 390.0 req_seconds_bucket{le="+Inf"} 390.0
Percentiles req_seconds_bucket{le="0.05"} 0.0 req_seconds_bucket{le="0.25"} 1.0 req_seconds_bucket{le="0.5"} 273.0 req_seconds_bucket{le="0.75"} 369.0 req_seconds_bucket{le="1.0"}
388.0 req_seconds_bucket{le="2.0"} 390.0 req_seconds_bucket{le="+Inf"} 390.0
Percentiles req_seconds_bucket{le="0.05"} 0.0 req_seconds_bucket{le="0.25"} 1.0 req_seconds_bucket{le="0.5"} 273.0 req_seconds_bucket{le="0.75"} 369.0 req_seconds_bucket{le="1.0"}
388.0 req_seconds_bucket{le="2.0"} 390.0 req_seconds_bucket{le="+Inf"} 390.0
None
Aggregation
Aggregation sum( rate( req_seconds_count[1m] ) )
Aggregation sum( rate( req_seconds_count[1m] ) )
Aggregation sum( rate( req_seconds_count[1m] ) )
Aggregation sum( rate( req_seconds_count[1m] ) )
Aggregation sum( rate( req_seconds_count{dc="west"}[1m] ) )
Aggregation sum( rate( req_seconds_count[1m] ) ) by (dc)
Percentiles histogram_quantile( 0.9, rate( req_seconds_bucket[10m] ))
Percentiles histogram_quantile( 0.9, rate( req_seconds_bucket[10m] ))
Percentiles histogram_quantile( 0.9, rate( req_seconds_bucket[10m] ))
Percentiles histogram_quantile( 0.9, rate( req_seconds_bucket[10m] ))
Percentiles histogram_quantile( 0.9, rate( req_seconds_bucket[10m] ))
None
None
Internal
Internal ❖ great for ad-hoc
Internal ❖ great for ad-hoc ❖ 1 expr per graph
Internal ❖ great for ad-hoc ❖ 1 expr per graph
❖ templating
PromDash
PromDash ❖ best integration
PromDash ❖ best integration ❖ former official
PromDash ❖ best integration ❖ former official ❖ now deprecated
❖ don’t bother
Grafana
Grafana ❖ pretty & powerful
Grafana ❖ pretty & powerful ❖ many integrations
Grafana ❖ pretty & powerful ❖ many integrations ❖ mix
and match!
Grafana ❖ pretty & powerful ❖ many integrations ❖ mix
and match! ❖ use this!
None
Alerts & Scrying
Alerts & Scrying ALERT DiskWillFillIn4Hours IF predict_linear( node_filesystem_free[1h], 4*3600) <
0 FOR 5m
Alerts & Scrying ALERT DiskWillFillIn4Hours IF predict_linear( node_filesystem_free[1h], 4*3600) <
0 FOR 5m
Alerts & Scrying ALERT DiskWillFillIn4Hours IF predict_linear( node_filesystem_free[1h], 4*3600) <
0 FOR 5m
Alerts & Scrying ALERT DiskWillFillIn4Hours IF predict_linear( node_filesystem_free[1h], 4*3600) <
0 FOR 5m
Alerts & Scrying ALERT DiskWillFillIn4Hours IF predict_linear( node_filesystem_free[1h], 4*3600) <
0 FOR 5m
Alerts & Scrying ALERT DiskWillFillIn4Hours IF predict_linear( node_filesystem_free[1h], 4*3600) <
0 FOR 5m
None
None
None
Environment
None
Apache nginx Django PostgreSQL MySQL MongoDB CouchDB redis Varnish etcd
Kubernetes Consul collectd HAProxy statsd graphite InfluxDB SNMP
Apache nginx Django PostgreSQL MySQL MongoDB CouchDB redis Varnish etcd
Kubernetes Consul collectd HAProxy statsd graphite InfluxDB SNMP
node_exporter
node_exporter cAdvisor
System Insight
System Insight ❖ load
System Insight ❖ load ❖ procs
System Insight ❖ load ❖ procs ❖ memory
System Insight ❖ load ❖ procs ❖ memory ❖ network
System Insight ❖ load ❖ procs ❖ memory ❖ network
❖ disk
System Insight ❖ load ❖ procs ❖ memory ❖ network
❖ disk ❖ I/O
mtail
mtail ❖ follow (log) files
mtail ❖ follow (log) files ❖ extract metrics using regex
mtail ❖ follow (log) files ❖ extract metrics using regex
❖ can be better than direct
Moar
Moar ❖ Edges: web servers/HAProxy
Moar ❖ Edges: web servers/HAProxy ❖ black box
Moar ❖ Edges: web servers/HAProxy ❖ black box ❖ databases
Moar ❖ Edges: web servers/HAProxy ❖ black box ❖ databases
❖ network
So Far
So Far ❖ system stats
So Far ❖ system stats ❖ outside look
So Far ❖ system stats ❖ outside look ❖ 3rd
party components
Code
cat-or.not
cat-or.not ❖ HTTP service
cat-or.not ❖ HTTP service ❖ upload picture
cat-or.not ❖ HTTP service ❖ upload picture ❖ meow!/nope meow!
from flask import Flask, g, request from cat_or_not import is_cat
app = Flask(__name__) @app.route("/analyze", methods=["POST"]) def analyze(): g.auth.check(request) return ("meow!" if is_cat(request.files["pic"]) else "nope!") if __name__ == "__main__": app.run()
from flask import Flask, g, request from cat_or_not import is_cat
app = Flask(__name__) @app.route("/analyze", methods=["POST"]) def analyze(): g.auth.check(request) return ("meow!" if is_cat(request.files["pic"]) else "nope!") if __name__ == "__main__": app.run()
from flask import Flask, g, request from cat_or_not import is_cat
app = Flask(__name__) @app.route("/analyze", methods=["POST"]) def analyze(): g.auth.check(request) return ("meow!" if is_cat(request.files["pic"]) else "nope!") if __name__ == "__main__": app.run()
pip install prometheus_client
from prometheus_client import \ start_http_server # … if __name__ ==
"__main__": start_http_server(8000) app.run()
process_virtual_memory_bytes 156393472.0 process_resident_memory_bytes 20480000.0 process_start_time_seconds 1460214325.21 process_cpu_seconds_total 0.169999999998 process_open_fds 8.0
process_max_fds 1024.0
process_virtual_memory_bytes 156393472.0 process_resident_memory_bytes 20480000.0 process_start_time_seconds 1460214325.21 process_cpu_seconds_total 0.169999999998 process_open_fds 8.0
process_max_fds 1024.0
process_virtual_memory_bytes 156393472.0 process_resident_memory_bytes 20480000.0 process_start_time_seconds 1460214325.21 process_cpu_seconds_total 0.169999999998 process_open_fds 8.0
process_max_fds 1024.0
process_virtual_memory_bytes 156393472.0 process_resident_memory_bytes 20480000.0 process_start_time_seconds 1460214325.21 process_cpu_seconds_total 0.169999999998 process_open_fds 8.0
process_max_fds 1024.0
process_virtual_memory_bytes 156393472.0 process_resident_memory_bytes 20480000.0 process_start_time_seconds 1460214325.21 process_cpu_seconds_total 0.169999999998 process_open_fds 8.0
process_max_fds 1024.0
process_virtual_memory_bytes 156393472.0 process_resident_memory_bytes 20480000.0 process_start_time_seconds 1460214325.21 process_cpu_seconds_total 0.169999999998 process_open_fds 8.0
process_max_fds 1024.0
None
from prometheus_client import \ Histogram, Gauge REQUEST_TIME = Histogram( "cat_or_not_request_seconds",
"Time spent in HTTP requests.")
from prometheus_client import \ Histogram, Gauge REQUEST_TIME = Histogram( "cat_or_not_request_seconds",
"Time spent in HTTP requests.") ANALYZE_TIME = Histogram( "cat_or_not_analyze_seconds", "Time spent analyzing pictures.")
from prometheus_client import \ Histogram, Gauge REQUEST_TIME = Histogram( "cat_or_not_request_seconds",
"Time spent in HTTP requests.") ANALYZE_TIME = Histogram( "cat_or_not_analyze_seconds", "Time spent analyzing pictures.") IN_PROGRESS = Gauge( "cat_or_not_in_progress_requests", "Number of requests in progress.")
@IN_PROGRESS.track_inprogress() @REQUEST_TIME.time() @app.route("/analyze", methods=["POST"]) def analyze(): g.auth.check(request) with ANALYZE_TIME.time(): result
= is_cat( request.files["pic"].stream) return "meow!" if result else "nope!"
@IN_PROGRESS.track_inprogress() @REQUEST_TIME.time() @app.route("/analyze", methods=["POST"]) def analyze(): g.auth.check(request) with ANALYZE_TIME.time(): result
= is_cat( request.files["pic"].stream) return "meow!" if result else "nope!"
AUTH_TIME = Histogram("auth_seconds", "Time spent authenticating.") AUTH_ERRS = Counter("auth_errors_total", "Errors
while authing.") AUTH_WRONG_CREDS = Counter("auth_wrong_creds_total", "Wrong credentials.") class Auth: # ... @AUTH_TIME.time() def auth(self, request): while True: try: return self._auth(request) except WrongCredsError: AUTH_WRONG_CREDS.inc() raise except Exception: AUTH_ERRS.inc()
AUTH_TIME = Histogram("auth_seconds", "Time spent authenticating.") AUTH_ERRS = Counter("auth_errors_total", "Errors
while authing.") AUTH_WRONG_CREDS = Counter("auth_wrong_creds_total", "Wrong credentials.") class Auth: # ... @AUTH_TIME.time() def auth(self, request): while True: try: return self._auth(request) except WrongCredsError: AUTH_WRONG_CREDS.inc() raise except Exception: AUTH_ERRS.inc()
AUTH_TIME = Histogram("auth_seconds", "Time spent authenticating.") AUTH_ERRS = Counter("auth_errors_total", "Errors
while authing.") AUTH_WRONG_CREDS = Counter("auth_wrong_creds_total", "Wrong credentials.") class Auth: # ... @AUTH_TIME.time() def auth(self, request): while True: try: return self._auth(request) except WrongCredsError: AUTH_WRONG_CREDS.inc() raise except Exception: AUTH_ERRS.inc()
AUTH_TIME = Histogram("auth_seconds", "Time spent authenticating.") AUTH_ERRS = Counter("auth_errors_total", "Errors
while authing.") AUTH_WRONG_CREDS = Counter("auth_wrong_creds_total", "Wrong credentials.") class Auth: # ... @AUTH_TIME.time() def auth(self, request): while True: try: return self._auth(request) except WrongCredsError: AUTH_WRONG_CREDS.inc() raise except Exception: AUTH_ERRS.inc()
@app.route("/analyze", methods=["POST"]) def analyze(): g.auth.check(request) with ANALYZE_TIME.time(): result = is_cat(
request.files["pic"].stream) return "meow!" if result else "nope!"
pip install prometheus_async
Wrapper from prometheus_async.aio import time @time(REQUEST_TIME) async def view(request): #
...
Goodies
Goodies ❖ aiohttp-based metrics export
Goodies ❖ aiohttp-based metrics export ❖ also in thread!
Goodies ❖ aiohttp-based metrics export ❖ also in thread! ❖
Consul Agent integration
Wrap Up
Wrap Up
Wrap Up ✓
Wrap Up ✓ ✓
Wrap Up ✓ ✓ ✓
ox.cx/p @hynek vrmd.de