The number appeared simple enough to understand without much discussion. Subtract the timestamp attached to a sports event from the time its record arrived, then describe the difference as delivery delay. In our 2022 review, that method produced an apparent median of 4.8 seconds across 600 event pairs from three feeds. The service sponsor wanted to use the figure in an infrastructure argument. Technical colleagues asked a prior question: did the clocks, event definitions and timestamp meanings support that comparison? The arithmetic was available before the evidence needed to interpret it. I had to decide whether the enterprise could use a persuasive number while that more basic question remained unresolved.
A subtraction needs a supported meaning
Two timestamps can look similar and describe different moments. One may be attached when an observation is recorded, another when a message is prepared or received. Their clocks also need a basis that makes comparison meaningful. I do not need to assume a particular software implementation to recognise the problem: a numerical difference cannot explain itself. The event pairing and timing context determine what the subtraction might measure. If those conditions are uncertain, adding decimal places or summarising many pairs does not remove the uncertainty. It can make an unsupported interpretation look more precise than the evidence allows.
The sponsor's concern about speed was legitimate. A live-data service needs to understand delivery performance, and users may care greatly about delays. I did not want the review to become a reason never to measure. The choice was between using the available figure as a whole-service account and doing the work needed to establish its basis. We committed ₹53 lakh to the timing review and diagnostic work. My team examined matching and interpretation with the relevant technical colleagues, who retained responsibility for the clocks and delivery systems. The data function's role was to ensure the account matched what those sources could support.
The review established a supported timing basis for 420 event pairs across two feeds. Their median was 2.3 seconds. The other 180 pairs belonged to a third feed whose clock evidence remained unresolved. That result was useful, but not in the way the initial presentation had suggested. It supplied a bounded measure for a defined subset and a clear question about the remainder. I withheld a whole-service latency claim. We did not fill the unresolved portion with the average from the other feeds or treat it as a set of inconvenient observations that could be removed without changing the meaning of the result.
The unresolved feed belongs in the account
The temptation to publish only the supported median was strong because it looked better than the earlier apparent figure. A reader might see 4.8 seconds followed by 2.3 seconds and assume the service had become faster. But the earlier measure was inadequately founded, and the later population differed. The two values did not form a valid before-and-after performance comparison. We had changed the account of the system, not demonstrated a change in the system's delivery. I kept the unresolved 180 pairs alongside the supported subset so that the limitation travelled with the number rather than remaining in an internal discussion.
This distinction is uncomfortable when a measurement review takes time and money. Colleagues understandably want to know what the investment achieved. The answer may be that the enterprise can now make a narrower and more defensible statement. That is a real improvement in its understanding, but it is not the same as an operating benefit. Converting it into a claim of faster delivery would lose the very discipline the review was meant to establish. The ₹53 lakh cost remained visible, as did the work still needed on the third feed. Neither was offset by a fictional performance gain inferred from the lower median.
I also wanted the next technical decision to start from the right question. If the evidence problem concerns clock meaning or event pairing, a general investment in faster infrastructure may not resolve it. That does not prove the infrastructure is adequate; it means the measurement cannot yet establish the proposed conclusion. Technical colleagues need room to investigate the actual source of uncertainty. A data leader can help by separating what is known about delivery from what is known about the record used to describe delivery. Those accounts may eventually connect, but they should not be treated as identical before the connection is supported.
Separate the measurement repair from the operating result
The broader lesson is to name the intervention accurately. A measurement repair changes definitions, evidence or handling so that the account becomes more faithful. An operating intervention changes the service itself. Both may be valuable and may occur together, but their effects need to be distinguished. A lower value after a method change does not automatically show that the operating intervention worked. In our case, I could explain the supported subset and the unresolved timing evidence. I could not claim an all-feed improvement or infer what the third feed's delay would have been under a valid measurement basis.
This way of thinking helps a board ask better questions without becoming a technical review team. Directors can ask whether a reported change reflects the business, the measurement method, the population or some combination. Management should be able to explain the material distinction in ordinary language. That does not require every clock detail in the board paper. It requires a clear statement of what the figure represents and why it supports the decision being requested. A metric that cannot survive that explanation should not acquire authority merely because it appears in a familiar dashboard beside other apparently settled numbers.
I now try to agree the intended timing claim before a collection is assembled for reporting. Which event marks the beginning? Which event marks the end? What evidence makes their times comparable? Which observations are included, and why? These are not procedural questions to ask only after an inconvenient result appears. They are part of designing a measure that can be used responsibly. The 420 supported pairs and the 180 unresolved ones gave us a more honest starting point. The achievement was understanding the boundary of the account, while leaving the service's actual performance to be established through evidence that could bear that claim.