<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Cihangir Celal Palacı on Medium]]></title>
        <description><![CDATA[Stories by Cihangir Celal Palacı on Medium]]></description>
        <link>https://medium.com/@cihangirpalaci?source=rss-64799645d1e7------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*5bSD_w_pbuftAMNma7J5rQ.jpeg</url>
            <title>Stories by Cihangir Celal Palacı on Medium</title>
            <link>https://medium.com/@cihangirpalaci?source=rss-64799645d1e7------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 04:58:36 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@cihangirpalaci/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Strengthening Data-Driven Decisions: Clickstream Anomaly Detection with Helyx]]></title>
            <link>https://medium.com/trendyol-tech/from-manual-checks-to-scalable-anomaly-detection-lessons-from-a-clickstream-platform-5f352373d337?source=rss-64799645d1e7------2</link>
            <guid isPermaLink="false">https://medium.com/p/5f352373d337</guid>
            <category><![CDATA[data-observability]]></category>
            <category><![CDATA[data-quality]]></category>
            <category><![CDATA[data-engineering]]></category>
            <category><![CDATA[anomaly-detection]]></category>
            <category><![CDATA[data-analytics]]></category>
            <dc:creator><![CDATA[Cihangir Celal Palacı]]></dc:creator>
            <pubDate>Wed, 04 Feb 2026 08:23:32 GMT</pubDate>
            <atom:updated>2026-02-04T08:27:14.903Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vBeqS5TPM0c_ibkOFN9RdQ.jpeg" /></figure><p>Modern product organizations run on clickstream data. Funnels, conversion rates, A/B tests, and growth metrics all assume one thing: <strong>the data reflects reality</strong>.</p><p>At Trendyol, clickstream data powers dozens of business-critical dashboards used daily by product, growth, and operations teams. Events generated across platforms, cultures, and screens are aggregated and transformed into signals that directly influence product roadmaps, experiments, and operational decisions.</p><p>For a long time, our biggest problem was not missing data. It was trusting data that looked perfectly fine, yet quietly told the wrong story.</p><h3>When Clickstream Data Breaks Quietly</h3><p>When a data pipeline crashes, the failure is obvious. Alerts fire, dashboards break, and fixes follow quickly. Clickstream data rarely fails this way.</p><p>Instead, it degrades silently. Events that should be emitted once get duplicated, parameters arrive partially or entirely null after releases, and event volumes drop only on specific platforms or cultures. Funnel ratios drift due to inconsistent client-side logic, often without any single metric clearly breaking.</p><p>Dashboards still load, and numbers still move. At first glance, everything looks fine, but the data may no longer fully represent reality. These edge-case degradations are often subtle and hard to catch early, yet they can quietly snowball if left unnoticed.</p><p>In a fast-moving environment like Trendyol, frequent mobile and web releases, temporary page closures, unexpected traffic spikes, and even fraud-related activity can all influence clickstream behavior. These changes do not always trigger obvious failures, but they can significantly distort event data without anyone realizing it.</p><h3>Why This Quiet Failure Impacts Business Decisions</h3><p>These issues are not just data quality problems. They directly affect how teams make decisions.</p><p>Product changes are evaluated using distorted signals. Experiments produce misleading results. Stakeholders gain confidence in dashboards that appear stable but no longer reflect real user behavior. In many cases, these problems surface weeks later during monthly reviews, long after decisions have already been made.</p><p>Manual SQL checks and static thresholds were not enough, not because they were wrong, but because they were inherently limited. There are patterns we do not yet know, thresholds we cannot predefine, and behavioral shifts that are impossible to track with human effort alone. At Trendyol scale, relying on visual inspection or fixed rules would require an unrealistic amount of manual work and still leave blind spots.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gQmtWsCtNHrJHuZLuhKT8Q.png" /><figcaption>A sudden and extreme deviation in click volume on the <em>search listing</em> screen. What looks like a healthy traffic surge at first glance is, in reality, a behavioral shift that requires investigation. This is exactly the kind of pattern that silently bypasses manual checks and static rules.</figcaption></figure><h3>From Manual Checks to a Trustworthy Detection System</h3><p>To address this challenge, the Data Reliability team at Trendyol built <strong>Helyx</strong>, our internal data observability platform. Helyx continuously monitors thousands of data products across the organization, executing data quality rules and anomaly detection. For a deeper dive into the architecture, design principles, and real-world impact of the platform, you can read more in <a href="https://medium.com/trendyol-tech/from-data-to-reliability-inside-helyx-our-data-observability-platform-b7e12cc043b0"><strong>Helyx: Data Observability Platform</strong></a>.</p><p>This article focuses on one of its core components: <strong>the clickstream anomaly detection system</strong>. Its role is simple but essential: detecting when clickstream data looks valid but behaves abnormally.</p><h3>Key Use Cases</h3><p>Instead of monitoring everything blindly, we focused our detection efforts on three high-impact areas:</p><p><strong>1. Click Count Anomaly Detection:</strong> We monitor event volumes across multiple dimensions (platform, culture, screen). Unexpected increases often indicate duplicated events or bot activity, while drops point to broken instrumentation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WLZjojLYUebKbKWEOJLRxQ.png" /><figcaption>A multi-country spike on the <em>my_account</em> screen revealed a newly introduced event that doubled expected volumes, while also helping teams identify unintended duplicate impressions early, and the duplication issue was fixed before propagating downstream.</figcaption></figure><p><strong>2. PID &amp; SID Anomaly Detection</strong></p><p>To support click count monitoring, we also track anomalies in <strong>PID–SID behavior</strong> within clickstream events. PID (Platform ID) and SID (Session ID) represent core identifiers of product interactions and user journeys, and their distributions remain relatively stable under normal conditions. Sudden shifts often indicate duplicated events, broken session tracking, or instrumentation regressions, even when overall event counts appear healthy.</p><p><strong>3. Null Percentage Monitoring:</strong> We track daily null ratios for selected parameters. Common causes include backend fields not being populated or type mismatches. These are dangerous because event counts may look healthy while the data itself becomes unusable.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IMHlahhemNMit3U_178G5Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Z51MhUdukDFfNulztkhHVw.png" /><figcaption>An example of a null percentage anomaly where click volumes appear healthy, but the underlying data quality degrades. The system detects an unexpected surge in null values for a critical parameter, enabling early investigation before the issue propagates to downstream analytics and decision-making.</figcaption></figure><h3>Learning Normal Behavior Without Static Thresholds</h3><p>Static rules like “alert if traffic drops 10%” do not scale. A small drop on a quiet weekday may be critical, while a large spike during a campaign may be expected.</p><p>By learning historical trends and seasonality per metric, the system understands what “normal” looks like for a specific hour on a specific day. This removes the need for manual threshold tuning and allows detection to adapt automatically as behavior evolves.</p><h3>A Voting-Based Ensemble Approach</h3><p>Early experiments showed that single-model anomaly detection was fragile. We operate over <strong>tens of thousands of heterogeneous metrics</strong>, each with different volatility, volume, and seasonality patterns. No single model could handle this diversity reliably.</p><p>Some models overreacted to trend changes, others missed seasonality, and many produced false positives that eroded trust. Instead of searching for a perfect model, we embraced a more realistic assumption: models will disagree, and that disagreement can be useful.</p><p>We adopted an ensemble strategy where multiple models evaluate the same signal independently. Each metric is analyzed by five complementary models:</p><ul><li>AutoARIMA</li><li>MSTL (Multiple Seasonal-Trend Decomposition using Loess)</li><li>DynamicOptimizedTheta</li><li>AutoETS (Error, Trend, Seasonality)</li><li>AutoTheta</li></ul><p>Each model covers a different failure mode: <strong>trend shifts, seasonality breaks, or variance explosions</strong>.</p><p>Rather than alerting on a single model’s output, we introduced a strict quorum rule. An anomaly is flagged only if <strong>at least four out of five models</strong> agree that the behavior deviates meaningfully from expectation. Signals flagged by fewer models are treated as noise.</p><p>This decision dramatically improved precision. When an alert fires, it is worth someone’s time.</p><h3>Architecture Overview</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*aSxlybOBzn7kYqigSx5c7g.png" /><figcaption>High-level architecture of the clickstream anomaly detection system. Source data flows from BigQuery into the ensemble models, and only high-confidence anomalies pass the voting layer to trigger actionable alerts.</figcaption></figure><p>The system follows a streamlined flow designed for early intervention:</p><ol><li><strong>Ingestion:</strong> Clickstream data is continuously ingested from BigQuery.</li><li><strong>Preprocessing:</strong> Raw events are transformed into time series aligned by metric, platform, and dimension.</li><li><strong>Model Evaluation:</strong> Each time series is independently evaluated by the ensemble of statistical models.</li><li><strong>Voting Layer:</strong> Only anomalies that pass the strict quorum rule are considered valid.</li><li><strong>Persistence &amp; Alerting:</strong> High-confidence anomalies are stored and routed to Slack for immediate visibility.</li><li><strong>Feedback Loop:</strong> User feedback is captured through status updates, turning each alert into organizational knowledge.</li></ol><h3>The Last Mile: From Detection to Action</h3><p>Detecting an anomaly is only half the battle. The other half is making sure the right person sees it, understands it, and acts on it. To avoid the bystander effect, we designed a two-layered consumption strategy: immediate triage via Slack and deeper investigation via the web UI.</p><h4>Smart Routing &amp; Visual Alerts (Slack)</h4><p>We don’t blast generic alerts to noisy channels. Each anomaly is routed using an ownership lookup map that identifies exactly which team owns the affected table or platform.</p><ul><li><strong>Targeted Tagging:</strong> The specific owner is tagged directly in the message thread, making responsibility clear.</li><li><strong>Self-Contained Context:</strong> Users don’t need to open a dashboard to understand the severity. Every alert includes a generated plot showing historical trend, expected range, and the deviation point.</li><li><strong>Quick Action:</strong> Users can categorize the anomaly (e.g., <em>Fraud</em>, <em>Incident</em>, <em>Expected</em>) immediately via an embedded dropdown.</li></ul><h4>Real Life Scenarios</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*IdMhcT38KAlYrNuLVI8pkw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*9KjUlM5SoH71FJ7v1JsxqA.png" /><figcaption>Sample Slack alerts: showing granular anomaly details and the magnitude of deviation, with Status dropdown menu embedded in the message</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gOXnx4pgTK9n-1IQAdl_rA.png" /><figcaption>A sample alert graph: The shaded area represents the expected range, and orange dots indicate anomalies</figcaption></figure><h4>The Anomaly Dashboard: Context and History</h4><p>While Slack is ideal for immediate reaction, many investigations require historical perspective. We built a dedicated interface where teams can explore long-term trends, compare patterns, and distinguish between recurring behavior and new issues.</p><p>The UI allows teams to:</p><ul><li>Filter anomalies by status, dimension, or date range</li><li>Review the full history of a metric to separate noise from real regressions</li><li>Update anomaly statuses in bulk, keeping the knowledge base clean and accurate</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LGoo543_qeD4AXIlVa7SfQ.png" /><figcaption>The Anomaly Management Hub: This dashboard acts as our organizational memory, allowing teams to filter historical anomalies by table, severity, or status (e.g., ‘Incident’). It turns transient alerts into a searchable knowledge base.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-g6Ynu8o5m9DdEWjsHeoig.png" /><figcaption>Anomaly Detail Page: Here, a user investigates a ‘Null Percentage’ anomaly. The timeline clearly visualizes the exact moment the field became 100% null, helping engineers correlate the data loss with specific deployment times.</figcaption></figure><h3>Closing the Loop: Models Detect, Humans Contextualize</h3><p>The model can say <em>“This looks weird,”</em> but it cannot say <em>“Why.”</em></p><p>Through the UI and Slack integration, domain experts classify anomalies with statuses such as:</p><ul><li><strong>Release Issue:</strong> A recent deployment caused the anomaly.</li><li><strong>Fraud/Bot Activity:</strong> Validated by the fraud team.</li><li><strong>Expected Behavior:</strong> A marketing campaign or feature flag change.</li></ul><p>This classification builds organizational memory. Over time, we don’t just know that something broke; we know <em>why</em> it broke, turning raw alerts into a valuable knowledge base.</p><h3>Lessons Learned: What Didn’t Work</h3><p>The system was not perfect from day one. We iterated through several failures before reaching stability:</p><ul><li><strong>Early Noise:</strong> Initial versions produced too many false positives until we implemented the strict voting mechanism.</li><li><strong>Data Instability:</strong> Some datasets never stabilized enough to be meaningful for ML models, teaching us that not every metric justifies model-level complexity.</li><li><strong>Context is King:</strong> We learned that models alone do not solve observability problems; context and feedback loops do.</li></ul><h3>Closing Thoughts</h3><p>Perfect anomaly detection does not exist. There will always be a trade-off between catching every single issue (recall) and avoiding alarm fatigue (precision).</p><p>By moving from manual, monthly discovery to a systematic, voting-based detection system, we didn’t just solve a data engineering problem. We solved a trust problem. We built a system that allows us to act within an actionable timeframe, ensuring that our decisions are based on reality, not noise.</p><p>Looking ahead, we are expanding the system with additional context signals to further reduce ambiguity and improve root cause analysis:</p><ul><li><strong>Application release versions:</strong> correlating anomalies with specific iOS, Android, and web deployments.</li><li><strong>Event creation times:</strong> distinguishing between delayed ingestion and true behavioral changes.</li><li><strong>Web traffic insights:</strong> incorporating traffic and routing information from web teams to better explain sudden drops or spikes.</li></ul><p>By enriching anomaly detection with release and traffic context, we aim to move from “something looks wrong” to “this changed because of X” even faster.</p><h3>About Us</h3><p>We’re building a team of the brightest minds in our industry. Interested in joining us? Visit the pages below to learn more about our open positions.</p><p><a href="https://careers.trendyol.com/tr">Home - Trendyol Careers</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5f352373d337" width="1" height="1" alt=""><hr><p><a href="https://medium.com/trendyol-tech/from-manual-checks-to-scalable-anomaly-detection-lessons-from-a-clickstream-platform-5f352373d337">Strengthening Data-Driven Decisions: Clickstream Anomaly Detection with Helyx</a> was originally published in <a href="https://medium.com/trendyol-tech">Trendyol Tech</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>