mirror of
https://github.com/prometheus/prometheus.git
synced 2026-05-05 04:16:15 +02:00
3bda9a3396
Some timestamp functions can be safely wrapped as a step invariant. Then once we do that we never need to check or unwrap at evaluation time. For instance; `timestamp(metric @ 1)` is step invariant, whereas `timestamp(abs(metric @ 1))` is not. Currently all `timestamp` `*parser.Call` are excluded from being considered step invariant since it is listed in the `AtModifierUnsafeFunctions` map. This PR adds an extra check for timestamp functions which consider the arguments, and if the argument is a simple VectorSelector then the entire `*parser.Call` can be safely wrapped as a step invariant. Signed-off-by: Andrew Hall <andrew.hall@grafana.com> Co-authored-by: Bryan Boreham <bjboreham@gmail.com>
310 lines
10 KiB
Plaintext
310 lines
10 KiB
Plaintext
load 10s
|
|
metric{job="1"} 0+1x1000
|
|
metric{job="2"} 0+2x1000
|
|
|
|
load 1ms
|
|
metric_ms 0+1x10000
|
|
|
|
# Instant vector selectors.
|
|
eval instant at 10s metric @ 100
|
|
metric{job="1"} 10
|
|
metric{job="2"} 20
|
|
|
|
eval instant at 10s metric @ 100s
|
|
metric{job="1"} 10
|
|
metric{job="2"} 20
|
|
|
|
eval instant at 10s metric @ 1m40s
|
|
metric{job="1"} 10
|
|
metric{job="2"} 20
|
|
|
|
eval instant at 10s metric @ 100 offset 50s
|
|
metric{job="1"} 5
|
|
metric{job="2"} 10
|
|
|
|
eval instant at 10s metric @ 100 offset 50
|
|
metric{job="1"} 5
|
|
metric{job="2"} 10
|
|
|
|
eval instant at 10s metric offset 50s @ 100
|
|
metric{job="1"} 5
|
|
metric{job="2"} 10
|
|
|
|
eval instant at 10s metric offset 50 @ 100
|
|
metric{job="1"} 5
|
|
metric{job="2"} 10
|
|
|
|
eval instant at 10s metric @ 0 offset -50s
|
|
metric{job="1"} 5
|
|
metric{job="2"} 10
|
|
|
|
eval instant at 10s metric @ 0 offset -50
|
|
metric{job="1"} 5
|
|
metric{job="2"} 10
|
|
|
|
eval instant at 10s metric offset -50s @ 0
|
|
metric{job="1"} 5
|
|
metric{job="2"} 10
|
|
|
|
eval instant at 10s metric offset -50 @ 0
|
|
metric{job="1"} 5
|
|
metric{job="2"} 10
|
|
|
|
eval instant at 10s metric @ 0 offset -50s
|
|
metric{job="1"} 5
|
|
metric{job="2"} 10
|
|
|
|
eval instant at 10s metric @ 0 offset -50
|
|
metric{job="1"} 5
|
|
metric{job="2"} 10
|
|
|
|
eval instant at 10s -metric @ 100
|
|
{job="1"} -10
|
|
{job="2"} -20
|
|
|
|
eval instant at 10s ---metric @ 100
|
|
{job="1"} -10
|
|
{job="2"} -20
|
|
|
|
# Millisecond precision.
|
|
eval instant at 100s metric_ms @ 1.234
|
|
metric_ms 1234
|
|
|
|
# Range vector selectors.
|
|
eval instant at 25s sum_over_time(metric{job="1"}[100s] @ 100)
|
|
{job="1"} 55
|
|
|
|
eval instant at 25s sum_over_time(metric{job="1"}[100s] @ 100 offset 50s)
|
|
{job="1"} 15
|
|
|
|
eval instant at 25s sum_over_time(metric{job="1"}[100s] offset 50s @ 100)
|
|
{job="1"} 15
|
|
|
|
eval instant at 25s sum_over_time(metric{job="1"}[100] @ 100 offset 50)
|
|
{job="1"} 15
|
|
|
|
eval instant at 25s sum_over_time(metric{job="1"}[100] offset 50s @ 100)
|
|
{job="1"} 15
|
|
|
|
# Different timestamps.
|
|
eval instant at 25s metric{job="1"} @ 50 + metric{job="1"} @ 100
|
|
{job="1"} 15
|
|
|
|
eval instant at 25s rate(metric{job="1"}[100s] @ 100) + label_replace(rate(metric{job="2"}[123s] @ 200), "job", "1", "", "")
|
|
{job="1"} 0.3
|
|
|
|
eval instant at 25s sum_over_time(metric{job="1"}[100s] @ 100) + label_replace(sum_over_time(metric{job="2"}[100s] @ 100), "job", "1", "", "")
|
|
{job="1"} 165
|
|
|
|
eval instant at 25s sum_over_time(metric{job="1"}[100] @ 100) + label_replace(sum_over_time(metric{job="2"}[100] @ 100), "job", "1", "", "")
|
|
{job="1"} 165
|
|
|
|
# Subqueries.
|
|
|
|
# 10*(1+2+...+9) + 10.
|
|
eval instant at 25s sum_over_time(metric{job="1"}[100s:1s] @ 100)
|
|
{job="1"} 460
|
|
|
|
# 10*(1+2+...+7) + 8.
|
|
eval instant at 25s sum_over_time(metric{job="1"}[100s:1s] @ 100 offset 20s)
|
|
{job="1"} 288
|
|
|
|
# 10*(1+2+...+7) + 8.
|
|
eval instant at 25s sum_over_time(metric{job="1"}[100s:1s] offset 20s @ 100)
|
|
{job="1"} 288
|
|
|
|
# 10*(1+2+...+7) + 8.
|
|
eval instant at 25s sum_over_time(metric{job="1"}[100:1] offset 20 @ 100)
|
|
{job="1"} 288
|
|
|
|
# Subquery with different timestamps.
|
|
|
|
# Since vector selector has timestamp, the result value does not depend on the timestamp of subqueries.
|
|
# Inner most sum=1+2+...+10=55.
|
|
# With [100s:25s] subquery, it's 55*4.
|
|
eval instant at 100s sum_over_time(sum_over_time(metric{job="1"}[100s] @ 100)[100s:25s] @ 50)
|
|
{job="1"} 220
|
|
|
|
# Nested subqueries with different timestamps on both.
|
|
|
|
# Since vector selector has timestamp, the result value does not depend on the timestamp of subqueries.
|
|
# Sum of innermost subquery is 220 as above. The outer subquery repeats it 3 times.
|
|
eval instant at 0s sum_over_time(sum_over_time(sum_over_time(metric{job="1"}[100s] @ 100)[100s:25s] @ 50)[3s:1s] @ 3000)
|
|
{job="1"} 660
|
|
|
|
# Testing the inner subquery timestamp since vector selector does not have @.
|
|
|
|
# Inner sum for subquery [100s:25s] @ 50 are
|
|
# at -50 nothing, at -25 nothing, at 0=0, at 25=2, at 50=5.
|
|
# This sum of 7 is repeated 3 times by outer subquery.
|
|
eval instant at 0s sum_over_time(sum_over_time(sum_over_time(metric{job="1"}[10s])[100s:25s] @ 50)[3s:1s] @ 200)
|
|
{job="1"} 21
|
|
|
|
# Inner sum for subquery [100s:25s] @ 200 are
|
|
# at 125=12, at 150=15, at 175=17, at 200=20.
|
|
# This sum of 64 is repeated 3 times by outer subquery.
|
|
eval instant at 0s sum_over_time(sum_over_time(sum_over_time(metric{job="1"}[10s])[100s:25s] @ 200)[3s:1s] @ 50)
|
|
{job="1"} 192
|
|
|
|
# Nested subqueries with timestamp only on outer subquery.
|
|
# Outer most subquery:
|
|
# at 925=360
|
|
# inner subquery: at 905=90+89, at 915=91+90
|
|
# at 950=372
|
|
# inner subquery: at 930=93+92, at 940=94+93
|
|
# at 975=380
|
|
# inner subquery: at 955=95+94, at 965=96+95
|
|
# at 1000=392
|
|
# inner subquery: at 980=98+97, at 990=99+98
|
|
eval instant at 0s sum_over_time(sum_over_time(sum_over_time(metric{job="1"}[20s])[20s:10s] offset 10s)[100s:25s] @ 1000)
|
|
{job="1"} 1504
|
|
|
|
# minute is counted on the value of the sample.
|
|
eval instant at 10s minute(metric @ 1500)
|
|
{job="1"} 2
|
|
{job="2"} 5
|
|
|
|
# timestamp() takes the time of the sample and not the evaluation time.
|
|
eval instant at 10m timestamp(metric{job="1"} @ 10)
|
|
{job="1"} 10
|
|
|
|
# The result of inner timestamp() will have the timestamp as the
|
|
# eval time, hence entire expression is not step invariant and depends on eval time.
|
|
eval instant at 10m timestamp(timestamp(metric{job="1"} @ 10))
|
|
{job="1"} 600
|
|
|
|
eval instant at 15m timestamp(timestamp(metric{job="1"} @ 10))
|
|
{job="1"} 900
|
|
|
|
# Time functions inside a subquery.
|
|
|
|
# minute is counted on the value of the sample.
|
|
eval instant at 0s sum_over_time(minute(metric @ 1500)[100s:10s])
|
|
{job="1"} 20
|
|
{job="2"} 50
|
|
|
|
# If nothing passed, minute() takes eval time.
|
|
# Here the eval time is determined by the subquery.
|
|
# [50m:1m] at 6000, i.e. 100m, is 50m to 100m.
|
|
# sum=51+52+...+59+0+1+2+...+40.
|
|
eval instant at 0s sum_over_time(minute()[50m:1m] @ 6000)
|
|
{} 1315
|
|
|
|
# sum=46+47+...+59+0+1+2+...+35.
|
|
eval instant at 0s sum_over_time(minute()[50m:1m] @ 6000 offset 5m)
|
|
{} 1365
|
|
|
|
# time() is the eval time which is determined by subquery here.
|
|
# 2901+...+3000 = (3000*3001 - 2899*2900)/2.
|
|
eval instant at 0s sum_over_time(vector(time())[100s:1s] @ 3000)
|
|
{} 295050
|
|
|
|
# 2301+...+2400 = (2400*2401 - 2299*2300)/2.
|
|
eval instant at 0s sum_over_time(vector(time())[100s:1s] @ 3000 offset 600s)
|
|
{} 235050
|
|
|
|
# timestamp() takes the time of the sample and not the evaluation time.
|
|
eval instant at 0s sum_over_time(timestamp(metric{job="1"} @ 10)[100s:10s] @ 3000)
|
|
{job="1"} 100
|
|
|
|
# The result of inner timestamp() will have the timestamp as the
|
|
# eval time, hence entire expression is not step invariant and depends on eval time.
|
|
# Here eval time is determined by the subquery.
|
|
eval instant at 0s sum_over_time(timestamp(timestamp(metric{job="1"} @ 999))[10s:1s] @ 10)
|
|
{job="1"} 55
|
|
|
|
|
|
clear
|
|
|
|
# Tests for @ modifier with empty data.
|
|
# Data only at 0s, 10s, 20s. Eval at timestamp with no data.
|
|
load 10s
|
|
up 1 2 3
|
|
|
|
# Functions that should return empty results when @ modifier points to timestamp with no data.
|
|
# These were panicking before the fix.
|
|
|
|
eval instant at 1111111s quantile_over_time(scalar(up) + 1, {__name__="up"}[1h:1m] @ 1111111)
|
|
|
|
eval instant at 1111111s predict_linear({__name__="up"}[1h:1m] @ 1111111, 0.1)
|
|
|
|
eval instant at 1111111s deriv({__name__="up"}[1h:1m] @ 1111111)
|
|
|
|
eval instant at 1111111s changes({__name__="up"}[1h:1m] @ 1111111)
|
|
|
|
eval instant at 1111111s resets({__name__="up"}[1h:1m] @ 1111111)
|
|
|
|
eval instant at 1111111s first_over_time({__name__="up"}[1h:1m] @ 1111111)
|
|
|
|
eval instant at 1111111s last_over_time({__name__="up"}[1h:1m] @ 1111111)
|
|
|
|
eval instant at 1111111s sum_over_time({__name__="up"}[1h:1m] @ 1111111)
|
|
|
|
eval instant at 1111111s avg_over_time({__name__="up"}[1h:1m] @ 1111111)
|
|
|
|
eval instant at 1111111s min_over_time({__name__="up"}[1h:1m] @ 1111111)
|
|
|
|
eval instant at 1111111s max_over_time({__name__="up"}[1h:1m] @ 1111111)
|
|
|
|
eval instant at 1111111s count_over_time({__name__="up"}[1h:1m] @ 1111111)
|
|
|
|
eval instant at 1111111s stddev_over_time({__name__="up"}[1h:1m] @ 1111111)
|
|
|
|
eval instant at 1111111s stdvar_over_time({__name__="up"}[1h:1m] @ 1111111)
|
|
|
|
eval instant at 1111111s mad_over_time({__name__="up"}[1h:1m] @ 1111111)
|
|
|
|
clear
|
|
|
|
# Additional tests specific to timestamp() and @ modifier usage.
|
|
load 10s
|
|
metric 0+1x10
|
|
metric_missing 0 _ 2 _ 4 _ 5 _ 6
|
|
|
|
# Return a vector where each sample is set to the metric value at T=11.
|
|
# Since T=11 falls within the [10s,20s) scrap window the sample value at T=10s is returned.
|
|
eval range from 0 to 60s step 10s metric @ 11
|
|
{__name__="metric"} 1 1 1 1 1 1 1
|
|
|
|
eval range from 0 to 60s step 10s abs(metric @ 11)
|
|
{} 1 1 1 1 1 1 1
|
|
|
|
# Return a vector where each sample's value is set to the timestamp of each sample in the given metric series
|
|
eval range from 0 to 60s step 10s timestamp(metric)
|
|
{} 0 10 20 30 40 50 60
|
|
|
|
# Return a vector where each sample's value is set to the timestamp for the metric's sample used at T=11s.
|
|
# The result is 10 since the metric at T=11s falls within the [10s,20s) scrape window.
|
|
# The result is the timestamp of the sample at T=10s
|
|
eval range from 0 to 60s step 10s timestamp(metric @ 11)
|
|
{} 10 10 10 10 10 10 10
|
|
|
|
# As above - illustrating the sample used at the upper end of the [10s,20s) scrape window.
|
|
eval range from 0 to 60s step 10s timestamp(metric @ 19)
|
|
{} 10 10 10 10 10 10 10
|
|
|
|
# As above - illustrating the transition to the next scrap window.
|
|
eval range from 0 to 60s step 10s timestamp(metric @ 20)
|
|
{} 20 20 20 20 20 20 20
|
|
|
|
eval range from 0 to 60s step 10s timestamp(metric_missing @ 0)
|
|
{} 0 0 0 0 0 0 0
|
|
|
|
# The timestamp of 0 is returned since the sample is missing from the [10s,20s) scrape window.
|
|
# As such, the previous sample from T=0s is returned.
|
|
eval range from 0 to 60s step 10s timestamp(metric_missing @ 10)
|
|
{} 0 0 0 0 0 0 0
|
|
|
|
eval range from 0 to 60s step 10s timestamp(metric_missing @ 20)
|
|
{} 20 20 20 20 20 20 20
|
|
|
|
# The timestamps for each step are returned since abs() returns a new vector with new [T,V] samples.
|
|
# Each sample in this vector has its value set to the absolute value of the sample value at T=10s, and its timestamp aligned to the step interval.
|
|
# This is unlike the above tests where timestamp() is operating on a vector with the original series samples.
|
|
eval range from 0 to 60s step 10s timestamp(abs(metric @ 11))
|
|
{} 0 10 20 30 40 50 60
|
|
|
|
eval range from 0 to 60s step 10s timestamp(abs(metric_missing @ 11))
|
|
{} 0 10 20 30 40 50 60
|
|
|
|
clear |