Late-Arriving Data

Late-arriving data is anything that reaches your reporting layer after the period it belongs to has already been reported. It is normal rather than exceptional, and almost every analytics stack has several sources of it at once. Mobile clients buffer events while offline and flush them when the connection returns. Failed deliveries are retried on a backoff that can span hours. Batch loads run once a day and land yesterday's rows this morning. Payments settle, refunds and chargebacks arrive weeks later and reverse revenue that was already counted. Attribution windows stay open for days, so a conversion keeps being credited to a click long after the click. The visible symptom is a chart that always dips at the right-hand edge, which teams learn to read past and then, once, read literally, usually in the meeting where the dip is presented as a decline. What makes it dangerous is not the movement but the asymmetry: the current period is always incomplete and every past period is closer to final, so any comparison of today against last week is comparing a partial number against a settled one. The discipline is to make completeness explicit rather than to chase it. Label recent periods as provisional in the report itself, and know roughly how long each source takes to settle so the label can come off on a schedule. When comparing periods, compare snapshots of the same age — what the previous month looked like three days in, against what this month looks like three days in — rather than a live figure against a closed one. Keep the arrival timestamp alongside the event timestamp, since without both you cannot tell a genuine change in behaviour from a change in when data showed up. And treat any number that has been sent outside the company as a figure that may need restating. The corollary for anyone buying software is that two vendors quoting the same metric may simply be waiting different lengths of time before they publish it, and the one with the more flattering number is often the one that waits longer.