FroggyAds

Measure the delay in a conversion report

· 4 min de lectura

An advertising team can see fewer accepted conversions at ten in the morning and recover most of the gap by lunch. To tell whether the report is slow or the campaign is changing, measure how long known events take to appear. This check needs the records from the system that accepts the business event, plus dated observations from the report that receives it. A daily total alone leaves both explanations open.

Choose the two timestamps

Start with one conversion action and one stable event identifier. For a lead campaign, use the time your lead system first marked a traceable lead as accepted. Use the first extraction time at which the same identifier appears in the receiving report. Convert both times to UTC and subtract the acceptance time from the first observed time. A click timestamp answers an attribution question and belongs in a separate column.

Take repeat extracts at a known interval and retain their timestamps. With fifteen minute extracts, the first visible extract gives an upper bound on the report’s arrival time. If an accepted lead was absent at 09:45 and appeared at 10:00, it arrived sometime in that interval. Keep the extraction cadence the same when comparing days. A switch from hourly to fifteen minute extracts can make the measured delay shorter even when the processing system has not changed.

Count the missing events alongside the delay

For each acceptance cohort, record how many unique events qualify and how many you have found by a fixed cutoff. Keep the unmatched identifiers available for investigation. If the cohort has fifteen eligible events and ten appear by the cutoff, report ten of fifteen observed with any delay summary. Assigning zero minutes to the other five would make the system appear faster. Excluding them without reporting the count would hide the remaining gap.

Use the same action, acceptance rule and observation window on both days. An overnight cohort should get the same amount of time to mature as a weekday morning cohort. Flag reversed timestamps for review. They may come from a clock conversion, an incorrect join, or a definition such as a timestamp that changes when an order is edited. The FroggyAds guide to reconciling conversion discrepancies gives the related checks for matching populations, accepted outcomes, report dates and processing delays.

Run a fixture before trusting the chart

Suppose ten matched events have observed delays of 3, 5, 8, 8, 11, 13, 20, 34, 55 and 120 minutes. The middle pair is 11 and 13, so a continuous median is 12 minutes. For a continuous ninetieth percentile the zero-based position is 0.90 × (10 − 1), or 8.1. The delay at position eight is 55 minutes and the next is 120. Interpolating ten percent of that interval gives 61.5 minutes. A nearest-rank ninetieth percentile would select 55 minutes from the same list.

Run those ten values through the production calculation and compare the output with the hand calculation. The PostgreSQL ordered-set aggregate reference documents percentile_cont and percentile_disc. For a separate check of the continuous result, use the linear interpolation option described in the pandas Series quantile reference. Keep fractional minutes through the calculation and round only when you present the result. Write down which definition the dashboard uses so someone can reproduce it in a different tool.

Put the numbers to work

Save both the count and the delay with every comparison. A small median alongside a rising ninetieth percentile suggests that a slower group deserves closer inspection. Break it down by the event source or ingestion route first. Keep the matched count next to each segment so that a percentile based on three observations cannot pass for a stable trend. Compare it with the rate of unmatched records, which can reveal a different kind of failure.

When a threshold is missed, open one fast record and one late record from the same cohort. Record the source acceptance time, the last extract in which the late record was absent and the first extract in which it was present. If the event is still absent at the agreed review time, send the exact identifier and cutoff to the person who manages that feed through your usual incident channel. Leave the advertising budget decision tied to completed data from the same eligible cohort.

This is a FroggyAds editorial example. The lead counts and timestamps are illustrative.