Stackrig computes its live numbers with a fast flow model. We check that model against a
slower, exact simulation that follows every request through the same designs. Below is
every check of the current engine: where the two agree within the tolerance, and the few
places where they do not, which we publish as known limits.
121 scenarios, 766 checks: 688 agree within the tolerance, 78 are published limits (explained below). No failures.
Tolerances: rates and utilization ±5 %, latency percentiles ±15 %, error shares ±0.5 points or ±15 %. Each value of the exact simulation is the mean of 8 independent runs.
Engine 0.1.0 · page build eefec0d (2026-10-14)
How we test
Every scenario below runs twice: once with the fast model that drives the live numbers in the playground, and once with a slow, exact simulation that follows every single request through the design.
The two must agree within ±5 % for throughput and utilization and within ±15 % for the latency percentiles; the exact simulation is repeated 8 times and averaged, so its own noise stays small.
Where they do not agree, we say so: the known limits below explain in plain words which situation it is (for example long overload, retries that feed on themselves, or a breaker right at its threshold) and in which direction the live numbers are off.
Agreeing with an exact simulation is not the same as agreeing with real machines, so we also run the same designs on real cloud VMs under load and publish how far the model is from those measurements, with the setup of every run.
Known limits
S7 · Overload with a finite buffer: With a full accept queue the live model’s p99 is about 20 % too high: the wait is nearly constant then, which the model’s distribution does not capture.
S8 · Timeouts and retries (stable): A retry right after a timeout meets the same congestion. The live model correlates retries through the shared wait and the orphaned work of timed-out calls, but still underrates long congestion episodes: the final error rate is about a fifth of the exact simulation’s.
S8b · Timeout below the healthy p99: When the retries plus the normal load are more than the service can handle, a retry storm is a second, self-sustaining state. The exact simulation falls into it after a random burst within seconds; the live model needs a trigger, then stays in it too (S37). The playground warns when a design is at risk of it.
S11 · Periodic traffic spikes (time averages): Over periodic spikes the p99 is about 18 % too high: transient queues are averaged, not replayed.
S16b Ca 2 · Bursty arrivals, 75 % busy: With very bursty arrivals the median is about 23 % too low; the tail (p95, p99) stays within tolerance.
S20b · Autoscaling step with cold starts (new instances 3× slower at first): A fresh instance is overloaded on its own while the live model averages its speed with the warm ones: the p99 is about 30 % too low and the autoscaler may decide one step higher.
S28b · Thundering herd: slow loads after a flush, with and without coalescing: Without coalescing and with the database deep in overload, the live model ends each load after the wait it met when it started, while in the exact simulation early loads of hot keys overtake the backlog: about 30–40 % too many misses reach the database. With coalescing it matches.
S33 · Circuit breaker against a dependency failing 60 %: it opens and the calls fail fast: While the breaker is open the dependency only sees the half-open probes, a few per cycle, and whether a cycle closes it is a coin toss: its arrival rate (under 1 per second here) is noisy in both models.
S33b · Circuit breaker near its threshold with fallback skip: it keeps opening and closing: Near its threshold a closed breaker reopens after exactly the minimum number of calls in the live model, while in the exact simulation a lucky window keeps it closed a little longer: the dependency sees about half as many calls (4 against 8 per second, of 200), and the callers about half as many errors (0.6 % against 1.1 %).
S35e · Saturated connection pool without a timeout: the caller turns the excess away: The pool is full and the caller’s threads all wait for it, so the caller turns the excess away. The exact simulation first queues it in the caller’s accept queue, the live model rejects it at once: error rate and throughput match, the p99 is about half.
S35f · Overloaded connection pool with a timeout and a retry: Overloaded with a timeout, the exact simulation tips into a collapse: once the database backs up, every waiting call gets a connection just before its timeout, so all calls reach the database and it stays overloaded. The live model stays in the state where the excess waits out its timeout. Both report the pool as full.
S36 · Blocking queue at 125 %: producers wait for space and the API turns the excess away: Saturated, a producer waits almost exactly the time the consumers need for the messages ahead. The live model spreads that wait wider: the median matches, the p99 is about a third too high.
S36b · Blocking queue with a 20 ms limit on the wait for space: With a 20 ms limit, most producers in the exact simulation wait close to it before they get space; the live model lets more of them in at once, so the median is lower. The tail and the error rate match.
S38h · Virtual users whose bottleneck is a read-only database of four equal servers: A few users spread over several equal database servers by chance, so at any moment some servers have a queue while others are idle. The live model spreads them evenly: its throughput is about 5 % too high and its p99 about a third too low.
S43b · Shared-core instance (GCP e2-small) at 85 % of the CPU it can sustain: On a shared core the requests that run at the same time split its limited speed. The live model gets the average right but underrates the slowest requests by about a fifth.
S44 · Several database reads per request, one after the other, through a connection pool, at 50 to 90 % load: With 4 reads through a pool of 2 at 70 % load the median is 15 % too low, at the edge of the tolerance. Every p95 and p99 agrees.
S44b · Several database reads per request on a database whose two threads share one core: With only 2 connections into a database whose threads share a core, the wait near full load moves a lot for a small change of the query time: the live model’s median and p95 are 15 to 18 % too low at 90 % load, and the median already from 70 % load with four reads per request. With 16 connections and four reads at 90 % load its p99 is 16 % below this run of the exact simulation, whose own p99 varies by a fifth between repeated runs there; against eight times as many runs it is 6 % low.
S46 · Very bursty arrivals (as measured for serverless functions) at one service, 50 to 90 % busy: With very bursty arrivals the slow requests are right (p95 within 11 %), but the typical request is 20 to 60 % too fast in the live model, and at half load its p99 is up to 40 % too high.
S46b · Very bursty arrivals at one service whose callers give up after 100 ms: the error rate: With very bursty arrivals and a short timeout the live model shows a quarter to a third too few failed requests at 75 % load and about a sixth too few at 90 %.
S46c · Very bursty arrivals at a service with four instances: With very bursty arrivals over four instances the slow requests are right (p95 and p99 within 10 %); the typical request is 15 to 35 % too fast in the live model.
S46d · Very bursty arrivals through a connection pool into a database: Behind a connection pool the p99 agrees from 75 % load. At half load it is a quarter to two fifths too high, as at a single service. The p95 at 75 % load is about a fifth too low, and the median is a quarter to three fifths too low: with bursts most requests wait longer than the live model says, while the slowest ones are right.
S47 n2 cv0.03 50 · Load balancer hands requests to 2 one-thread servers in turn: near-constant work, 50 % busy: A server that gets every n-th request and does near-constant work hardly queues, and the little waiting that is left is more even than the live model assumes: its p99 is up to about a quarter too high. The median matches.
S47 n2 cv0.03 70 · Load balancer hands requests to 2 one-thread servers in turn: near-constant work, 70 % busy: A server that gets every n-th request and does near-constant work hardly queues, and the little waiting that is left is more even than the live model assumes: its p99 is up to about a quarter too high. The median matches.
S47 n2 cv0.03 90 · Load balancer hands requests to 2 one-thread servers in turn: near-constant work, 90 % busy: A server that gets every n-th request and does near-constant work hardly queues, and the little waiting that is left is more even than the live model assumes: its p99 is up to about a quarter too high. The median matches.
S47 n2 cv2 50 · Load balancer hands requests to 2 one-thread servers in turn: very uneven work, 50 % busy: Work this uneven is outside the range the live model is validated for: its median is up to about a third too high, its p99 up to about a quarter.
S47 n2 cv2 70 · Load balancer hands requests to 2 one-thread servers in turn: very uneven work, 70 % busy: Work this uneven is outside the range the live model is validated for: its median is up to about a third too high, its p99 up to about a quarter.
S47 n2 cv2 90 · Load balancer hands requests to 2 one-thread servers in turn: very uneven work, 90 % busy: Work this uneven is outside the range the live model is validated for: its median is up to about a third too high, its p99 up to about a quarter.
S47 n4 cv0.03 70 · Load balancer hands requests to 4 one-thread servers in turn: near-constant work, 70 % busy: A server that gets every n-th request and does near-constant work hardly queues, and the little waiting that is left is more even than the live model assumes: its p99 is up to about a quarter too high. The median matches.
S47 n4 cv2 50 · Load balancer hands requests to 4 one-thread servers in turn: very uneven work, 50 % busy: Work this uneven is outside the range the live model is validated for: its median is up to about a third too high, its p99 up to about a quarter.
S47 n4 cv2 70 · Load balancer hands requests to 4 one-thread servers in turn: very uneven work, 70 % busy: Work this uneven is outside the range the live model is validated for: its median is up to about a third too high, its p99 up to about a quarter.
S47 n4 cv2 90 · Load balancer hands requests to 4 one-thread servers in turn: very uneven work, 90 % busy: Work this uneven is outside the range the live model is validated for: its median is up to about a third too high, its p99 up to about a quarter.
S47 n8 cv0.03 90 · Load balancer hands requests to 8 one-thread servers in turn: near-constant work, 90 % busy: A server that gets every n-th request and does near-constant work hardly queues, and the little waiting that is left is more even than the live model assumes: its p99 is up to about a quarter too high. The median matches.
S47 n8 cv2 50 · Load balancer hands requests to 8 one-thread servers in turn: very uneven work, 50 % busy: Work this uneven is outside the range the live model is validated for: its median is up to about a third too high, its p99 up to about a quarter.
S47 n8 cv2 70 · Load balancer hands requests to 8 one-thread servers in turn: very uneven work, 70 % busy: Work this uneven is outside the range the live model is validated for: its median is up to about a third too high, its p99 up to about a quarter.
S47 n8 cv2 90 · Load balancer hands requests to 8 one-thread servers in turn: very uneven work, 90 % busy: Work this uneven is outside the range the live model is validated for: its median is up to about a third too high, its p99 up to about a quarter.
Timeouts while a backlog grows: When a server falls behind for so long that requests wait out their timeout, the live model shows the failures up to one timeout too early: when the request that will time out arrives, not when its timeout runs out. Example: 2,000 requests a second into a server that finishes 1,600, with a timeout of 30 s. The live model shows the first failures after about 117 s; the exact simulation and the arithmetic say 150 s. From then on both agree.
A retry budget in a burst: A retry budget saves up the retries it did not use and can spend them when a burst begins. The live model saves up nothing: in the busiest one to two seconds of a burst it shows about a third to a half fewer calls than the exact simulation, so the load on the service behind it looks lower than it is for those seconds. After that both agree.
A database disk near its limit: Commits that arrive while the disk is flushing share the next flush. Up to 90 % of the disk’s throughput the live model’s wait for a flush is within about a tenth of a commit-by-commit simulation, and its p99 within about a seventh. Above that its p99 runs high: at 95 % by about a fifth.
Real-cloud measurements
The same designs, run on real cloud VMs and measured under load (as of 2026-09-30). Setup: the service and, where named, PostgreSQL each on its own Google Cloud n2-standard-2 VM (2 vCPU), the load generator on a third VM; the load rises in steps of 60 s, at a constant rate or with random (Poisson) arrivals. The table shows how far the live model is from the measurement on average, over the load steps up to 90 % CPU.
Mean absolute deviation of the live model from real machines
Run
Arrivals
Throughput
p50
p95
p99
S1: one CPU-bound service (2 vCPU, n2)
constant
0.0 %
6.6 %
6.7 %
6.3 %
S1: one CPU-bound service (2 vCPU, n2)
Poisson
1.0 %
22.6 %
23.1 %
26.9 %
S2: service → Postgres (2 vCPU each, n2)
constant
0.0 %
130.6 %
97.6 %
44.0 %
S2: service → Postgres (2 vCPU each, n2)
Poisson
0.2 %
99.8 %
30.7 %
42.1 %
Throughput matches the real machines within 1 %. For a CPU-bound service the latency percentiles are within 7 % at a constant rate and within about 25 % under random (Poisson) arrivals. For a service in front of PostgreSQL the real requests take under a millisecond, so small fixed costs dominate and the model’s median is up to about 1.7× off. Near the database’s capacity the model still underestimates the slowest requests, because requests there also wait for a free database connection in the application, which we are still calibrating. Until then, read Stackrig’s tails near a database’s capacity as optimistic.
Every check
Showing 766 of 766 checks.
S1 · One service, 4 threads, 75 % busy
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
s
rate
300
300
−0.0 %
±5.0 %
within
s
util
0.750
0.750
+0.1 %
±5.0 %
within
end-to-end
p50
11.84
11.60
+2.1 %
±15.0 %
within
end-to-end
p95
41.50
40.53
+2.4 %
±15.0 %
within
end-to-end
p99
63.48
63.48
0.0 %
±15.0 %
within
S2 · Load balancer → service → database with a replica
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
db
rate
1210
1209
+0.1 %
±5.0 %
within
s
util
0.550
0.550
−0.0 %
±5.0 %
within
db
util
0.682
0.680
+0.3 %
±5.0 %
within
end-to-end
p50
16.36
15.87
+3.1 %
±15.0 %
within
end-to-end
p95
39.55
36.62
+8.0 %
±15.0 %
within
end-to-end
p99
55.18
53.22
+3.7 %
±15.0 %
within
S3 · Cache in front of the database
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
db
rate
400
401
−0.4 %
±5.0 %
within
db
util
0.700
0.704
−0.6 %
±5.0 %
within
end-to-end
p50
2.11
2.11
0.0 %
±15.0 %
within
end-to-end
p95
11.60
11.60
0.0 %
±15.0 %
within
end-to-end
p99
22.22
22.22
0.0 %
±15.0 %
within
S4 · Parallel and sequential fan-out
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
db
rate
600
603
−0.4 %
±5.0 %
within
db
util
0.600
0.603
−0.5 %
±5.0 %
within
end-to-end
p50
35.64
35.64
0.0 %
±15.0 %
within
end-to-end
p95
65.43
65.43
0.0 %
±15.0 %
within
end-to-end
p99
86.91
86.91
0.0 %
±15.0 %
within
S5 · Queue with workers
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
w
rate
199
199
+0.1 %
±5.0 %
within
w
util
0.745
0.743
+0.3 %
±5.0 %
within
q
qp99
58.11
55.18
+5.3 %
±15.0 %
within
w
np99
73.24
73.24
0.0 %
±15.0 %
within
S6 · Rate limit at 1.2× the traffic
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
c
err
0.167
0.168
−0.7 %
±15.0 %
within
end-to-end
p50
1.79
1.79
0.0 %
±15.0 %
within
end-to-end
p99
5.31
5.43
−2.2 %
±15.0 %
within
S7 · Overload with a finite buffer
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
c
err
0.169
0.169
−0.2 %
±15.0 %
within
s
util
0.997
0.998
−0.1 %
±5.0 %
within
end-to-end
p50
69.34
79.10
−12.3 %
±15.0 %
within
end-to-end
p95
143
127
+12.3 %
±15.0 %
within
end-to-end
p99
193
158
+22.2 %
±15.0 %
limit
S7c · Overload with a finite buffer (constant arrivals)
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
c
ok per s from t = 0
399
399
−0.1 %
±5.0 %
within
S8 · Timeouts and retries (stable)
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
db
rate
268
270
−1.0 %
±5.0 %
within
s-db-read
edgeErr
0.0671
0.0793
−15.3 %
±15.0 %
limit
c
err
0.00117
0.00643
−81.8 %
±15.0 %
limit
end-to-end
p50
4.46
4.46
0.0 %
±15.0 %
within
end-to-end
p95
12.08
12.33
−2.0 %
±15.0 %
within
end-to-end
p99
18.80
18.31
+2.7 %
±15.0 %
within
S8b · Timeout below the healthy p99
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
db
rate
290
750
−61.4 %
±5.0 %
limit
c
err
0.0591
1.00
−94.1 %
±15.0 %
limit
S9 · Thread-pool exhaustion
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
s
slots
0.892
0.891
+0.1 %
±5.0 %
within
c
err
0.00
0.00
0.0 %
±15.0 %
within
end-to-end
p50
110
110
0.0 %
±15.0 %
within
end-to-end
p95
236
232
+1.7 %
±15.0 %
within
end-to-end
p99
316
309
+2.5 %
±15.0 %
within
S10 · Highly variable service times
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
s
util
0.800
0.799
+0.1 %
±5.0 %
within
end-to-end
p50
24.66
22.22
+11.0 %
±15.0 %
within
end-to-end
p95
127
124
+2.4 %
±15.0 %
within
end-to-end
p99
213
197
+7.9 %
±15.0 %
within
S11 · Periodic traffic spikes (time averages)
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
s
util
0.636
0.636
−0.0 %
±5.0 %
within
end-to-end
p50
12.33
12.33
0.0 %
±15.0 %
within
end-to-end
p99
146
118
+24.0 %
±15.0 %
limit
S12 · The “Web app with database” template from the live demo
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
db
util
0.440
0.440
−0.0 %
±5.0 %
within
api
util
0.187
0.187
+0.2 %
±5.0 %
within
end-to-end
p50
21.24
21.24
0.0 %
±15.0 %
within
end-to-end
p95
55.18
55.18
0.0 %
±15.0 %
within
end-to-end
p99
86.91
86.91
0.0 %
±15.0 %
within
S13 · Slot-bound queue consumer
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
w
rate
60.00
60.02
−0.0 %
±5.0 %
within
w
slots
0.825
0.826
−0.1 %
±5.0 %
within
w
util
0.0750
0.0752
−0.2 %
±5.0 %
within
q
qp99
197
197
0.0 %
±15.0 %
within
S14 · Writes into a read-through cache
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
k
rate
2250
2254
−0.2 %
±5.0 %
within
db
rate
300
300
−0.2 %
±5.0 %
within
db
util
0.525
0.525
−0.1 %
±5.0 %
within
k
util
0.0225
0.0225
−0.1 %
±5.0 %
within
k
np99
14.28
14.04
+1.7 %
±15.0 %
within
end-to-end
p50
3.45
3.39
+1.8 %
±15.0 %
within
end-to-end
p95
10.13
9.89
+2.5 %
±15.0 %
within
end-to-end
p99
17.33
17.33
0.0 %
±15.0 %
within
S15 · 8 threads on 4 cores with hyperthreads, 80 % busy
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
s
rate
1050
1052
−0.2 %
±5.0 %
within
s
util
0.761
0.761
−0.1 %
±5.0 %
within
end-to-end
p50
11.60
11.35
+2.2 %
±15.0 %
within
end-to-end
p95
38.57
39.55
−2.5 %
±15.0 %
within
end-to-end
p99
62.01
63.48
−2.3 %
±15.0 %
within
S15b · 2 vCPUs on 1 core, constant arrivals
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
s
rate
230
230
−0.0 %
±5.0 %
within
s
util
0.698
0.698
−0.0 %
±5.0 %
within
end-to-end
p50
6.53
6.53
0.0 %
±15.0 %
within
end-to-end
p95
7.26
7.39
−1.7 %
±15.0 %
within
end-to-end
p99
7.63
7.75
−1.6 %
±15.0 %
within
S16a Ca 0.5 · Smoother-than-random arrivals, 75 % busy
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
s
rate
300
300
−0.1 %
±5.0 %
within
s
util
0.750
0.749
+0.1 %
±5.0 %
within
end-to-end
p50
10.38
10.13
+2.4 %
±15.0 %
within
end-to-end
p95
22.22
21.24
+4.6 %
±15.0 %
within
end-to-end
p99
30.52
28.56
+6.8 %
±15.0 %
within
S16b Ca 2 · Bursty arrivals, 75 % busy
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
s
rate
300
299
+0.5 %
±5.0 %
within
s
util
0.750
0.747
+0.4 %
±5.0 %
within
end-to-end
p50
14.04
18.31
−23.3 %
±15.0 %
limit
end-to-end
p95
54.20
56.15
−3.5 %
±15.0 %
within
end-to-end
p99
83.01
83.01
0.0 %
±15.0 %
within
S16c Ca 0 chain · Constant arrivals through service → database, 80 % busy
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
s
util
0.300
0.300
−0.0 %
±5.0 %
within
db
util
0.800
0.800
−0.0 %
±5.0 %
within
db
np99
13.55
12.57
+7.8 %
±15.0 %
within
end-to-end
p50
8.18
8.42
−2.9 %
±15.0 %
within
end-to-end
p95
14.28
13.55
+5.4 %
±15.0 %
within
end-to-end
p99
18.80
16.85
+11.6 %
±15.0 %
within
S17 · Timeout storm without retries
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
c (arrivalCv 1)
goodput per s from t = 0
138
155
−10.5 %
±15.0 %
within
c (arrivalCv 0)
goodput per s from t = 0
145
144
+0.8 %
±15.0 %
within
S17b · Finite queue whose full-buffer wait is about the timeout
Where
Metric
Live model
Exact simulation
Deviation
Tolerance
Result
s
util
1.00
1.00
+0.0 %
±5.0 %
within
c
err
0.458
0.521
−12.2 %
±15.0 %
within
c-s
edgeErr
0.458
0.521
−12.2 %
±15.0 %
within
S18 · Network round trip in front of a tight timeout, 1 retry