{"id":4840,"date":"2026-11-20T07:35:00","date_gmt":"2026-11-20T07:35:00","guid":{"rendered":"https:\/\/dataopsschool.com\/blog\/data-report-five-whys-steps\/"},"modified":"2026-11-20T07:35:00","modified_gmt":"2026-11-20T07:35:00","slug":"data-report-five-whys-steps","status":"publish","type":"post","link":"https:\/\/dataopsschool.com\/blog\/data-report-five-whys-steps\/","title":{"rendered":"How to Trace a Broken Data Report With Five Whys, Step by Step"},"content":{"rendered":"<p>The weekly revenue table said one thing, the finance export said another, and the analyst who owned the pipeline had a sinking feeling. The analyst had already patched this discrepancy twice, each time by deleting duplicate rows by hand. This how-to follows that analyst through a five-step version of asking why, because the method is easier to learn when you see it applied to a data engineering problem from beginning to end. It is a composite scenario, kept deliberately generic so you can map it to your own stack.<\/p>\n<h2>Step 1: Write the symptom as a measurable statement<\/h2>\n<p>&#8220;The numbers look off&#8221; cannot be investigated. The analyst rewrote it as: &#8220;Total orders in the reporting table exceed the source count by roughly the same small share every Monday.&#8221; That sentence has a place, a rhythm and a direction. Those three details immediately narrow the search, because a problem that appears on Mondays points to something weekly, such as a batch export, a time zone boundary or a backfill.<\/p>\n<p>Do this first, before anyone suggests a fix. If you cannot state the symptom with a where and a when, you risk solving a neighbouring problem and leaving the real one untouched.<\/p>\n<h2>Step 2: Ask the first why and write the answer as a fact<\/h2>\n<p>Why did the table exceed the source count? Because some order rows appeared twice. This answer is checkable, and the analyst checked it with a grouped count on the order identifier. It held. A good first answer is almost always a symptom restated at a lower level, and that is fine. It gives the second question something solid to stand on.<\/p>\n<p>Resist the temptation to jump to the deletion script here. The script would turn the answer into a chore, and chores have a way of becoming permanent.<\/p>\n<h2>Step 3: Keep asking, and let each answer come from evidence<\/h2>\n<p>The next questions went like this, each answered by looking at something, never by guessing.<\/p>\n<ol>\n<li><strong>Why did rows appear twice?<\/strong> The Monday load ingested a file that overlapped with the previous Sunday&#8217;s file.<\/li>\n<li><strong>Why did the files overlap?<\/strong> The source system re-sent the final hours of each week to cover late-arriving records.<\/li>\n<li><strong>Why did the load accept overlapping data?<\/strong> The ingestion step appends rows and has no rule for recognising a record it has already seen.<\/li>\n<li><strong>Why was there no such rule?<\/strong> The original design assumed each file would be disjoint, and that assumption was never recorded anywhere except the builder&#8217;s memory.<\/li>\n<\/ol>\n<p>Four answers in, the story has moved from &#8220;bad numbers&#8221; to &#8220;an unwritten assumption&#8221;. That is the moment a chain becomes valuable, because assumptions that live in one head are exactly the sort of root cause that a team can write down, test and fix.<\/p>\n<h2>Step 4: Decide whether you have reached a cause you can change<\/h2>\n<p>Ask yourself three quick questions. Can the team alter this? Would altering it have prevented the original symptom? Does it explain why the problem repeated? The analyst answered yes to all three for &#8220;the design assumed disjoint files&#8221;. A fifth why added a final layer: why was the assumption never written down? Because the onboarding document for the pipeline lists tables and schedules but has no section for assumptions about inputs.<\/p>\n<p>That is the cause carried into the planning meeting. It was not &#8220;someone built it wrong&#8221;, which would have been unfair and unhelpful. It was &#8220;our documentation template has no place for input assumptions&#8221;, which is a gap anyone can close.<\/p>\n<h2>Step 5: Make one change and set a date for a follow-up look<\/h2>\n<p>The fix was a pair of small things: an idempotent load keyed on the order identifier plus a version stamp, and a new section in the pipeline document titled &#8220;What we assume about inputs&#8221;. The analyst also booked a ten-minute check four Mondays later to confirm the discrepancy had stopped. A check that small is easy to keep, and it turns a hopeful fix into a verified one. Once the numbers matched, the note was closed. When a similar mismatch appeared in a different table, the new section prompted a colleague to ask about overlapping files in the first hour instead of the third week.<\/p>\n<h2>What the story is really about<\/h2>\n<p>It is tempting to read this as a tale about deduplication. It is really a tale about where the investigation chose to stop. A team that stops at the second answer ends up with a cleanup script. A team that goes to the fourth or fifth ends up with a design note and a safer pipeline. Same people, same data, different ending. The only thing that changed was a willingness to ask one extra question, and then another, before reaching for the quick patch. Keep this up for twelve months and a team ends up holding a quiet library of written assumptions that makes every new hire productive sooner.<\/p>\n<h2>Practical tips for the first time you try this<\/h2>\n<ul>\n<li>Keep a notebook open and write each answer on its own line. Seeing the chain on a page makes weak links obvious.<\/li>\n<li>Invite the person who knows the system best, but make the newest team member ask the questions. Fresh eyes ask the &#8220;obvious&#8221; things that veterans skip.<\/li>\n<li>Use words like &#8220;because&#8221; and &#8220;which meant&#8221; so that each line really connects to the one above it.<\/li>\n<li>If you find yourself at an answer that begins &#8220;somebody should have&#8221;, go back one line and ask what process would have made the right action automatic.<\/li>\n<\/ul>\n<h2>A note on borrowing structure from elsewhere<\/h2>\n<p>Data people are good at spotting well-organised catalogues. A medical-tourism style directory such as <a href=\"https:\/\/bestorthohospitals.com\/\">a site that lays out hospital, doctor, blog, forum and procedure sections side by side<\/a> makes a tidy illustration of a taxonomy, and taxonomy is precisely what a lineage diagram needs: every table and field has an obvious home, so a missing or duplicated one is visible quickly.<\/p>\n<p>If you want tools that catch problems like our overlapping files earlier, two pieces in our archive are worth a look. One covers <a href=\"https:\/\/dataopsschool.com\/blog\/how-monte-carlo-stops-bad-data-before-it-hits-production\/\">how a data observability product blocks poor-quality inputs before production<\/a>, and the other is a practical guide to <a href=\"https:\/\/dataopsschool.com\/blog\/great-expectations-for-data-validation-in-dataops-a-practical-guide\/\">writing validation expectations for your datasets<\/a>. Neither replaces the questioning habit, but both make good places to put the answer once you have it.<\/p>\n<h2>Closing the loop<\/h2>\n<p>When the next discrepancy turns up, start with the measurable statement, collect your answers from the data rather than from memory, and stop only at something your team can change and own. Then write it down where newcomers will find it. A report that stays right is a small thing, but it is built from exactly this kind of patient, repeated question.<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Duplicate rows kept returning every Monday. This step-by-step story shows how five whys moved a data analyst from the bad total to the unwritten assumption behind it.<\/p>\n","protected":false},"author":6,"featured_media":4841,"comment_status":"","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[814],"tags":[],"class_list":["post-4840","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-dataops"],"_links":{"self":[{"href":"https:\/\/dataopsschool.com\/blog\/wp-json\/wp\/v2\/posts\/4840","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/dataopsschool.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/dataopsschool.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/dataopsschool.com\/blog\/wp-json\/wp\/v2\/users\/6"}],"replies":[{"embeddable":true,"href":"https:\/\/dataopsschool.com\/blog\/wp-json\/wp\/v2\/comments?post=4840"}],"version-history":[{"count":0,"href":"https:\/\/dataopsschool.com\/blog\/wp-json\/wp\/v2\/posts\/4840\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/dataopsschool.com\/blog\/wp-json\/wp\/v2\/media\/4841"}],"wp:attachment":[{"href":"https:\/\/dataopsschool.com\/blog\/wp-json\/wp\/v2\/media?parent=4840"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/dataopsschool.com\/blog\/wp-json\/wp\/v2\/categories?post=4840"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/dataopsschool.com\/blog\/wp-json\/wp\/v2\/tags?post=4840"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}