Skip to content

News

The documented limit: find the constraint before you buy capacity

November 19, 2026 · PenDraco

PenDraco reads a documented limit as the report of one test on one configuration. In one engagement, a new design carried an embedded analytics platform past its documented limit to about 90,000 concurrent users. Measure the constraint before you buy capacity.

In 1967 Gene Amdahl estimated what a second processor buys. With 2 processors that share memory, he wrote, "one would find approximately 2.2 times as much hardware". The return was smaller: "The resulting performance achieved would be about 1.8". By his estimate, more than double the hardware buys less than double the performance.

An owner or a chief operating officer meets that arithmetic when a system reaches the limit in the vendor's documentation. One proposal buys capacity. The other finds the real constraint first. The public sources favor the second, within limits stated below.

A documented limit reports a test

A capacity figure has a date and a configuration. Google's Site Reliability Engineering book shows why. It gives this example: "a cluster of X machines could handle Y queries per second three months ago". The book asks if the cluster still can, "given changes to the system". Its advice is to "Use load testing rather than tradition".

Amazon Web Services lists anti-patterns for load tests. The first is "Performing load testing on deployments that are not the same configuration as your production". RFC 2544 is a benchmarking method for network devices. Section 7 requires each report of results to include the software version and the exact configuration.

Together the 3 texts tie a capacity figure to the setup and the date of its test. PenDraco's reading takes one more step. A documented limit is a figure of that kind. It describes the design that the vendor tested, and the platform may allow others. No source states that step.

Added hardware can return less

Amdahl concluded that parallel effort pays only when the sequential part gains nearly as much. Neil Gunther's universal scalability law, as his 2008 paper states it, adds a second parameter. One parameter covers work that waits in line, or "serialization or queueing". The other covers "inconsistent copies of data". When the second is present, capacity has a peak. Beyond the peak "the throughput becomes retrograde", Gunther writes. Each added unit lowers the total.

So a hardware purchase can miss in 2 ways. The new capacity can wait in the same line as the old. It can also push the system past its peak. The Site Reliability Engineering book reports a cost that request counts miss. With many low-traffic clients, "health checking on the connections can require more resources" than the requests.

One design went past its documented limit

One engagement shows that a documented limit can move. For Cisco WebEx, PenDraco's founder designed a new scaling architecture for an embedded analytics platform. The design carried the platform past its documented limit to about 90,000 concurrent users.

The founder is Steve A. Onoja, Founder and Principal Engineer of PenDraco. He did the work during his decade at Qlik, his employer at the time. The sources cited here describe the practice in general. This article gives no further detail of that engagement.

Identity is a cost that the design sets

Public documents price one design choice in detail: how a platform checks its users. Before a business system serves a request, it resolves who the user is and what the user can see. The user's session presents a token as proof. The Internet Engineering Task Force describes 2 designs for the check. Its documents call the identity service the authorization server. In RFC 7662, the application asks that server about a token "using an HTTP POST". Each check is a network call. The RFC allows a cache "at the cost of liveness of the information". Liveness means freshness. In RFC 9068, the answers travel inside a signed token. The application checks it "without further authorization server involvement".

Microsoft documents the same choice for an application that acquires tokens from its identity platform. A token from the cache costs "one cache access". A token from the service costs "two cache accesses + one HTTP call". The same page gives example times of "~100 ms versus ~700 ms". A second page covers the application that asks too often. It receives an HTTP 429 response and no tokens. "Often, a 429 response indicates the application isn't caching and reusing tokens correctly", the page says.

PenDraco reads the 2 groups together. The first ties a capacity figure to a tested configuration. The second prices one design choice: how often a platform asks the identity service about a user. PenDraco's inference is that this choice belongs to the configuration behind a documented limit. More application servers do not lower its cost, because each server is another caller of the same service. So a change in the identity design can move a documented limit. Added hardware may not. No source in either group says so.

What the sources do not show

No source calls a documented limit a tested configuration. The 3 texts in the first section come nearest, and none speaks of a platform's published limit. RFC 2544 covers network devices, and it does "not specify an Internet standard of any kind". The step from those texts to a vendor's limit is PenDraco's inference.

No source says that identity is often the constraint, and none gives a frequency. The engagement above is 1 case, and this article does not state its constraint. The sources document a load cost for 2 designs only. One calls the identity service on each request. The other acquires a new token on each request. RFC 9068 still requires a signature check on every token, and no source calls that check a limit.

The design with fewer calls has a price. RFC 7662 warns that a cached answer can outlive a revoked token. So the identity design is a security decision as well as a capacity decision.

Gunther's model counts processors. Its reach to user load rests on his statement that "any conclusions also hold for software scalability". Amdahl's figures are estimates, and Microsoft's 2 times are examples without test conditions. The Site Reliability Engineering book reports Google's practice without a data set. Capacity is sometimes the right purchase. With idle resources, the book says, "adding tasks can be the most expedient way to recover from the outage".

Measurement comes before the purchase order

The decision is an order of work: measure the constraint, then buy for it. These 4 steps come first.

  1. Ask the vendor for the configuration behind the documented limit. Ask for the test date and the identity setup as well.
  2. Check every resource for utilization and saturation before you price hardware. Brendan Gregg's USE Method gives the checklist, and it extends to "software imposed limits".
  3. Load test the production configuration until a component breaks. The Site Reliability Engineering book warns: "You won't know in advance which component may hit the wall first".
  4. Count the calls to the identity service behind one user request. Then price each design in load and in freshness.

A capacity problem with another cause is in The nightly job became nightly.

If your system sits at a documented limit, a PenDraco engineer measures the constraint inside your environment. The embedded sprint puts one workflow into production in 30 to 60 days. Get in touch.

Sources

← Back to news