<?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 Angel One on Medium]]></title>
        <description><![CDATA[Stories by Angel One on Medium]]></description>
        <link>https://medium.com/@angelone?source=rss-b06a776e75f4------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*1Wo-SkCExuOxC433bGCiMA.png</url>
            <title>Stories by Angel One on Medium</title>
            <link>https://medium.com/@angelone?source=rss-b06a776e75f4------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 04:25:42 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@angelone/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[Seeing performance the way our users see it: building Real User Monitoring]]></title>
            <link>https://medium.com/angel-tech/seeing-performance-the-way-our-users-see-it-building-real-user-monitoring-c4f250341acb?source=rss-b06a776e75f4------2</link>
            <guid isPermaLink="false">https://medium.com/p/c4f250341acb</guid>
            <dc:creator><![CDATA[Angel One]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 04:42:22 GMT</pubDate>
            <atom:updated>2026-11-11T04:42:22.751Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Introduction</strong></p><p>Our server dashboards were green, and our error rates were low. Yet, a persistent blind spot remained: frontend performance — what the user sees — was completely dark. We lacked visibility into the issues users were facing — issues that happened sometimes on a specific screen, sometimes on a specific phone, sometimes only after a release. The gap between “the backend is healthy” and “the user had a bad time” is where most mobile performance problems live. Closing it meant we had to start measuring from the device outward, not the server in.</p><p><strong>Problem Statement</strong></p><p><strong>Problem 1: The old pipe wasn’t built for this</strong></p><p>As our user base grew, generic application logs quickly showed their limitations. Standard logs simply lack the granularity to answer complex performance questions. While a log can isolate a specific application error, it cannot measure screen load dynamics, session fluidity, or the exact environmental constraints of the end user.</p><p>Trying to compress this rich context into generic log formats stripped away the very structure that makes performance analysis effective. To identify regressions, we require native, indexable fields for device, screen, and API context. Additionally, our existing log infrastructure was built for point-search troubleshooting, not for handling massive write volumes and scanning millions of events across wide time horizons.</p><p><strong>Problem 2: Server health ≠ user experience</strong></p><p>A backend can be fast and correct, and the user can still have a bad time. The app might retry silently, a screen might render before its data arrives, or a specific phone model might choke on an animation the backend never sees. Server-side logs and APMs measure server-side truth. They cannot see the extra few milliseconds of client-side queueing, the frame that froze for a second, or the button a user tapped four times before giving up. We lacked a mechanism to measure the experience from inside the session that actually happened, on the device that actually ran it.</p><p><strong>Problem 3: Detection speed mattered as much as detection</strong></p><p>A performance regression that takes a week to surface translates to seven days of a degraded user experience and silent churn. We needed the capability to detect localised regressions while they were still contained — such as a build that runs 10% slower on a specific device tier, or an API boundary that begins timing out for a single cellular carrier — before they escalated into systemic outages or support queues.</p><p>Achieving this meant moving away from historical batch reporting. Relying on daily or weekly batch processing meant we could only audit what went wrong last month, leaving us blind to a fundamental operational question: “<em>Is anything broken right now?</em>”</p><p><strong>Problem 4: Aggregates hide regressions per release</strong></p><p>Every release is a bet that a new build improves upon the last. However, evaluating these deployments solely through global, historical aggregates creates a massive blind spot. Because overall fleet metrics blend multiple versions together, a localized regression in a newly canary-released build is mathematically averaged away by the stable performance of the rest of the fleet.</p><p>Solving these four problems meant we couldn’t simply extend our existing logging pipeline; we needed a purpose-built path for client telemetry from device to dashboard. Before designing the underlying system infrastructure, we first had to define the exact client-side signals required to eliminate our operational blind spots.</p><p><strong>What We Capture</strong></p><p>Before designing the underlying system, we identified a handful of critical client-side signals that do most of the heavy lifting. We categorise these into three primary pillars:</p><p><strong>Core Performance &amp; UX</strong></p><ul><li><strong>Screen load time:</strong> Tracks phase-by-phase load waterfalls per screen.</li><li><strong>Frame smoothness &amp; jank rate:</strong> Aggregates on-device frame timings to isolate UI stutter.</li><li><strong>Rage &amp; dead clicks:</strong> Monitors repeated taps or non-responsive areas as proxies for silent user frustration.</li></ul><p><strong>Network &amp; API Health</strong></p><ul><li><strong>API error rates &amp; latency:</strong> Measures client-observed performance across multi-percentile distributions (p50, p95, and p99).</li><li><strong>API traffic profiles:</strong> Analyses request/response volume, payload sizes, and status-code distribution over time.</li><li><strong>Network handshake &amp; socket performance:</strong> Monitors app-to-service connection dynamics, including DNS lookup latency, TCP/SSL handshake times, and new vs. reused connection ratios.</li></ul><p><strong>Environment &amp; Deployment Context</strong></p><ul><li><strong>Release &amp; build adoption:</strong> Tracks session shares by build version to gate build-over-build comparisons.</li><li><strong>Cohort standardisation:</strong> Normalises data by device model, network generation, and carrier populations to isolate localised hardware or network degradations.</li></ul><p>Everything below covers the architecture required to scale this — how we ingest, process, and transform these raw client-side signals from an on-device interaction into a live, low-latency dashboard query.</p><p><strong>Architecture</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/731/1*BOLW9AofhVgOD8waZ6V-oQ.png" /></figure><p><strong>On-Device Aggregation &amp; Batching</strong></p><p>To prevent telemetry from exhausting device resources, the app aggregates signals on-device first. It captures a screen’s load phases, a window of frame timings, or a burst of taps in one spot — instead of shipping a separate event for every single frame or tap.</p><p>Smoothness is a perfect example of this philosophy. On Android, we leverage <a href="https://developer.android.com/topic/performance/jankstats">Jetpack’s JankStats library</a>. JankStats tracks and flushes a window of dozens of frames at once, delivering an aggregated jank rate for that specific window rather than one event per frame. Furthermore, that window isn’t sent over the network the moment it is ready. Instead, it is cached on-device alongside other session events and shipped later as a single, unified batch. This represents a deliberate cost and performance trade-off: we exchange single-frame resolution for fewer, richer events over the wire, successfully avoiding a performance-killing firehose of raw samples.</p><p><strong>The Data Pipeline Stack</strong></p><p>Once shipped, these batches land on a dedicated telemetry ingestion API, kicking off a specialised, real-time data pipeline.</p><ul><li><strong>Ingestion &amp; Streaming:</strong> Telemetry data flows directly from the API into an <a href="https://aws.amazon.com/msk/">Amazon MSK</a> cluster to handle high-throughput write volumes safely.</li><li><strong>Processing</strong>: A Kafka consumer pipeline built on <a href="https://vector.dev/">Vector</a> picks up the data, processes it, and routes it to its destinations.</li><li><strong>Hot Storage:</strong> The primary destination is <a href="https://clickhouse.com/">ClickHouse</a>, a columnar data store optimized for real-time aggregation. We enforce a 30-day retention Time-to-Live (TTL) on this tier, keeping our highly indexable hot storage lean and performant for all live dashboard queries.</li><li><strong>Cold Storage:</strong> Simultaneously, a durable copy of the data is streamed into <a href="https://aws.amazon.com/s3/">Amazon S3</a> to serve as our cold tier. This layer maintains a 3-year retention window, providing a cost-effective environment for deep historical trend audits and seasonal baseline analysis.</li></ul><p><strong>Scale, Tied to the Market</strong></p><p>Because our traffic patterns directly mirror market activity, our pipeline must accommodate extreme throughput fluctuations. In the quiet early morning hours, the system idles at a baseline of just a few hundred gigabytes per hour. However, when the market hits the opening bell, the system absorbs an intense surge, scaling from a few hundred gigabytes to several terabytes of data streaming into Kafka within a single hour at up to hundreds of thousands of requests per second — representing a massive spike of more than 20x.</p><p>Across a full trading day, the pipeline smoothly processes tens of terabytes and tens of billions of individual events. We built the architecture to handle these predictable spikes seamlessly within a fixed-capacity Kafka cluster, completely eliminating the need for manual resizing or dynamic infrastructure adjustments between market hours.</p><p>The API stays just as fast at either extreme: ingest latency holds in the low double-digit milliseconds at the 95th percentile, and the average event stays under a single kilobyte, whether at baseline volume or peak terabyte-scale throughput.</p><p><strong>Data Modelling &amp; OTel Alignment</strong></p><p>At the data layer, every event uses a flat, standardized schema. Each payload carries a metric name, a numerical value, a unit of measurement, a timestamp, and a flat attribute map. <br>This architecture is conceptually very close to the <a href="https://opentelemetry.io/">OpenTelemetry</a> (OTel) metric data model. While we do not claim full OTel-wire-format compliance today, we consider our design “OTel-aligned” rather than strictly “OTel.”</p><p>A sample event payload for illustration:</p><p>{</p><p>“metric_name”: “screen_view.load_time”,</p><p>“value”: xxx,</p><p>“unit”: “ms”,</p><p>“timestamp”: “2026–08–28T19:54:00Z”,</p><p>“attributes”: {</p><p>“device.model”: “Pixel 8”,</p><p>“device.os”: “Android 14”,</p><p>“app.version”: “v2.4.0”,</p><p>“screen.name”: “Screen-1”</p><p>}</p><p>}</p><p><strong>The Dashboard — A Tour</strong></p><p><em>Panels in this section are re-rendered with synthetic data to illustrate dashboard structure and reasoning, not production volumes.</em></p><p><strong>Client-Side API Error Rate, Latency &amp; Trends</strong></p><p>The client experiences an API call the exact same way the user does. It captures everything a server-only view completely misses: DNS resolution failures, TLS handshake issues, flaky networks, and client-side timeouts that the server never logged because the request simply never arrived.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*McX5C7uTy4cbTP_ygA5AdA.png" /></figure><p>We isolate HTTP 4xx and 5xx errors into dedicated, real-time views and trend these metrics over a two-week window, comparing current data directly against the same windows from one and two weeks prior. Annotated tooltips provide point-in-time cross-comparison. This ensures that when an engineer asks, “<em>Is today worse than normal?</em>” they have an actual historical baseline for comparison, rather than just a raw, contextless number.</p><p>A granular data table provides a real-time snapshot of network performance, grouped by endpoint. While the time-series charts above flag <em>when</em> an anomaly occurred, this table isolates <em>where</em> it is happening. By pairing call volume and unique user counts with client/server error splits and multi-percentile latency distributions (<strong>p50, p95, and p99</strong>), engineers can instantly identify if a spike is a widespread platform failure or a localised endpoint degradation impacting the users</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lVPZ05Zy9M_5EQXJlrQ5Dw.png" /></figure><p><strong>Release Adoption</strong></p><p>Before attempting to compare performance across two different application versions, we must verify that both builds have accumulated enough real-world traffic to be statistically significant.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CamSKAdXoYRxLcgqKt0k-w.png" /></figure><p>This dashboard view answers a foundational question: “<em>How far has this specific build actually rolled out?</em>” We measure this by session share percentage rather than raw session counts, ensuring that low-volume outliers do not skew our broader performance metrics. This panel effectively serves as a strict rollout gate — a newly deployed build with only a handful of active sessions does not get to qualify as a valid data point yet.</p><h3>HTTP Status Mix Over Time</h3><p>Not every failure is built the same way. To understand the true nature of an issue, we split all inbound traffic into stacked bands over time: successes, redirects, client errors (4xx), and server errors (5xx).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/913/1*phhzGgELERSvpQZLAeaJ5w.png" /></figure><p>Visualising these categories as a stacked trend makes a shift in the “shape” of our traffic immediately obvious long before a raw, aggregated error-rate number moves enough to trigger a page. For example, an expanding client-error band right after a release usually signals an API contract mismatch or an authentication change. Conversely, any sudden appearance of a server-error band points squarely at the backend infrastructure, vindicating the client application instantly.</p><h3>Rage-click Monitoring</h3><p>A rage click — defined as several fast, consecutive taps in the same small area — is a user explicitly telling us, <em>“This should have worked, and it didn’t.”</em> It is the ultimate metric for silent frustration, capturing user friction long before anyone bothers to file a support ticket.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/912/1*3OtMtia6jalMkLmZNGiMMw.png" /></figure><p>Our dashboard surfaces exactly which screens attract the highest volume of these frantic interactions. To maintain strict data privacy, we aggregate this telemetry generalised entirely to the <strong>screen level</strong> rather than mapping the specific UI control or coordinate a user clicked. This approach gives engineering teams exactly enough macro-visibility to prioritise UX fixes, without collecting unnecessary telemetry that could reconstruct exactly what any single person was doing during their session.</p><h3>Device, Network, Carrier — Standardised</h3><p>Naively averaging performance data by device model or network type is a statistical trap. For instance, a high-end device whose owners disproportionately open one resource-heavy screen will always look “slow” on paper, regardless of how powerful its hardware is.</p><p>To bypass this noise, our dashboard normalises the data. Instead of presenting raw latency averages, this view asks a much narrower, smarter question: <em>“Does this specific population experience more slow or failed sessions than an average device performing the exact same actions?”</em></p><p>A score that spikes above our calculated baseline indicates a genuine, localised engineering issue for that specific cohort — proving that the app is underperforming on that hardware, rather than the users simply interacting with the app differently. We track this normalised variance across three core axes:</p><p>· <strong>Device Model:</strong> Spotting screen load degradations on specific hardware footprints.</p><p>· <strong>Network Generation:</strong> Isolating how the app behaves on 3G vs. 4G vs. 5G connections.</p><p>· <strong>Carrier:</strong> Detecting regional telecommunication routing anomalies or CDN blind spots.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/990/1*6fx3MQbPCR-ZFOHwoQoEZw.png" /></figure><p>For each axis, we intentionally surface only the headline operational metrics: <em>“Is this population performing worse than expected?”</em> and <em>“By how much?”</em></p><h3>Catching a Bad Release</h3><p>The clearest real-world application of all this telemetry is catching an unstable release before it escalates into a customer support queue.</p><p>This view compares our core performance vitals — such as screen load times, session smoothness, and API error rates — on a strict build-over-build basis. Crucially, these comparisons are gated by our release adoption rules to ensure that a newly deployed version has accumulated enough sessions to make the data statistically trustworthy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Cp1c9Q6C4YnyQZ1BuWUrOg.png" /></figure><p>When a backwards-compatibility break or a critical regression occurs, it leaves a very specific footprint: a sharp spike in errors or latency restricted entirely to a single build number, while all historical versions remain flat. This distinct signature is exactly how our team differentiates between a localised deployment failure (<em>“this specific build broke something”</em>) and a macro systemic issue (<em>“our entire infrastructure is having a bad day”</em>).</p><h3>Conclusion</h3><p>None of this replaces server-side observability — it sits right next to it.</p><p>A healthy backend is still the foundation of a reliable application, and traditional server logs and APMs remain critical for debugging core infrastructure. What our RUM pipeline adds is the vital perspective that server logs structurally cannot have: the session exactly as the device experienced it, from the initial tap to the final render.</p><p>Investing in a dedicated client-side telemetry path has completely transformed how we handle mobile reliability. By executing on this shift — moving our perspective from the device outward rather than the server in — we have turned a major operational blind spot into an early-warning system. We no longer wait for anomalies to cause widespread user frustration or bubble up through external support channels. We now possess the real-time, normalised data needed to diagnose edge-case slowdowns, validate adoption thresholds, and isolate environment-specific degradations instantly.</p><p>Ultimately, performance engineering is about validating the actual experience of the person holding the screen. By bridging the gap between server truth and client reality, we have built a defensive engineering gate — the exact perspective that catches a bad release before it becomes a bad user experience.</p><p>About the Author:</p><p>Soundar Rajan L</p><p>Software Development Engineer 3 — Application Development</p><p><em>Angel One Limited; SEBI Regn. No: INZ000161534</em></p><p><em>Disclaimer: Investments in securities market are subject to market risks, read all the related documents carefully before investing </em><a href="https://bit.ly/4u4zuXN"><em>https://bit.ly/4u4zuXN</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c4f250341acb" width="1" height="1" alt=""><hr><p><a href="https://medium.com/angel-tech/seeing-performance-the-way-our-users-see-it-building-real-user-monitoring-c4f250341acb">Seeing performance the way our users see it: building Real User Monitoring</a> was originally published in <a href="https://medium.com/angel-tech">Angel One Square</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How Angel One does A/B Experimentations at Scale]]></title>
            <link>https://medium.com/angel-tech/how-angel-one-does-a-b-experimentations-at-scale-0b10d3a367bd?source=rss-b06a776e75f4------2</link>
            <guid isPermaLink="false">https://medium.com/p/0b10d3a367bd</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ab-ex]]></category>
            <category><![CDATA[fintech]]></category>
            <dc:creator><![CDATA[Angel One]]></dc:creator>
            <pubDate>Mon, 26 Oct 2026 04:36:40 GMT</pubDate>
            <atom:updated>2026-10-30T09:41:49.474Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>A/B experimentation sits at the core of decision making when we roll out new features. We serve millions of customers every day, and each feature impacts user experience at scale. For that we need more than the gut feeling of Product and Business teams to decide what reaches the end user. We let you on the journey on how we build our Inhouse Experimentation Platform in series of blogs, while this one will address the basics of Experimentation. Experimentation is not as new as we might assume. Agriculture, Pharma, and many other industries have used it as standard practice for decades.</p><p><strong>What is A/B testing ?</strong></p><p>A/B testing is a statistical method that splits users into two or more groups and analyses their data to determine which performs best. Most of the time there are two groups: control (default behaviour) and test (new feature), but there also exist variants like A/A tests and multivariate testing and supported in the framework. <br> <br>To run these experiments reliably we need a system that adheres to the following <strong>functional requirements</strong>:</p><ul><li>Should assign variants <strong>consistently, </strong>a user assigned the test variant should remain in test as long as the config remains the same.</li><li>Should distribute users <strong>as close to the expected distribution</strong> as possible.</li><li>Should do this <strong>cheaply</strong>, compute-wise.</li><li>Should be able to run <strong>multiple non-overlapping experiments</strong>.</li></ul><p>We revisit each of these in “How We Meet These Requirements.” Section.</p><p><strong>Architecture</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/940/1*2wET0gh1vZpLM9hLHet8ew.png" /></figure><p>We have mainly two types of clients to serve</p><ol><li>Client Devices (Android, iOS, Web)</li><li>Microservices</li></ol><p>We put significant effort into making evaluation <strong>locally executable</strong>, so that compute stays distributed and scales easily. This matters for three reasons:</p><ul><li><strong>Latency</strong> — no network hop sits on the hot path of a request.</li><li><strong>Availability</strong> — evaluation keeps working even if the central service is degraded.</li><li><strong>Scale and cost</strong> — evaluation compute is spread across every caller instead of concentrated in one service.</li></ul><p>There is one deliberate split:</p><ul><li><strong>Microservices</strong> evaluate <strong>locally</strong>. Each Experimentation SDK calls the central Experimentation Service at regular intervals to fetch experiment metadata, then evaluates its own.</li><li><strong>Client Devices</strong> evaluate <strong>remotely</strong> via an Eval API call to the Experimentation Service and cache the data for current trading window.</li></ul><p><strong>How does evaluation happen?</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*A3sa7OP0CHUnCOn5mleaZQ.png" /></figure><p>All gates are pretty much self-explanatory, except Mutex, that deserves its own section <a href="https://angelbrokingpl.atlassian.net/wiki/spaces/BB/pages/5330960426/How+AngelOne+does+A+B+Experimentations+at+Scale#Mutex">Mutex</a>.</p><p><strong>Mutex</strong></p><p>A Mutex can be configured with up to 100 quotas. Each experiment receives a disjoint slice (for e.g. Exp1: 33, Exp2: 33, Exp3: 34). The gate hashes each user into a quota and assigns them to the experiment that owns it. Because slices do not overlap, a user joins at most one experiment, so experiments cannot contaminate each other’s results. This lets many experiments run concurrently: mutually exclusive ones share a Mutex, while independent ones use separate Mutexes and run in parallel.</p><p><strong>Dry Run</strong></p><p>(for e.g. 10k users visit your page):</p><p># Experiment Setup</p><p>- Experiments: Exp1,Exp2,Exp3</p><p>- Mutex: 100 Quotas</p><p>- Distribution: Exp1(33 quota), Exp2(33 quota), Exp3 (34 quota)</p><p><strong>Evaluation:</strong></p><ul><li>At the <strong>Mutex Gate</strong>, 33% of them (3300) get exposed to Exp1.</li><li>At the <strong>Exposure Gate</strong>, 50% of them (1650) enter since audience percentage = 50.</li><li>At the <strong>Cohort Gate</strong>, the cohort condition is checked; assuming 1/2 qualify, 833 remain.</li><li>At the <strong>Variant Gate</strong>, variants are assigned, and each variant (control and test) gets ~50% of users.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/945/1*RQ4e9pZy_tbmSKejlj2-9w.png" /></figure><p><strong>How do we choose users at each gate?</strong></p><p>Of the four gates, only the <strong>Cohort Gate</strong> has a definite lookup: does the user have the cohort or not — a simple element search in an array. The other three gates need logic to decide whether to pass a user while respecting each gate’s distribution policy (e.g. the Exposure Gate’s 50% audience percentage).</p><p>For that we use <strong>hashing</strong>. At each hashing gate we compute:</p><p>bucket = hash(seed, userId) % 100</p><p>This maps every user to a number from 0 to 99. A gate with a 50% policy passes users whose bucket is less than 50. The same mechanism runs at the Mutex, Exposure, and Variant gates.</p><p>The <strong>quality</strong> of this hash — how uniformly it spreads users across buckets — is what determines whether our distributions are trustworthy.</p><p><strong>How We Meet These Requirements</strong></p><p>Let’s revisit the functional requirements we listed earlier.</p><p>Should assign variants consistently — a user assigned the test variant should remain in test as long as the config remains the same.</p><p>This property is called <strong>Sticky Experiment</strong>. To solve it, the entire flow must be deterministic. Three of the four gates are hashing-based, and the fourth (Cohort) is a metadata check that is already deterministic. The hashing gates become deterministic by fixing the seed at each layer.</p><p>The Mutex entity stores 1 seed; the Experiment entity stores 2 seeds (one for the Exposure Gate, one for the Variant Gate). These are fixed on creation and never change during the entity’s lifecycle. Because bucket = hash(seed, userId) % 100, a fixed seed means the same user always lands in the same bucket.</p><p>Precisely: a user’s variant is stable <strong>iff</strong> (userId, seeds, gate thresholds, quota allocation) are unchanged. Changing an unrelated config field does not move a user; changing a seed, a threshold, or the quota boundaries can.</p><p>Should distribute users as close to the expected distribution as possible.</p><p>This property is called <strong>Hashing Uniformity</strong>. To solve it, we ran multiple rounds of tests across multiple hashing functions, measuring the deviation of the per-bucket distribution from the expected uniform spread, and picked the function with the least skew.</p><p>Should be able to do this cheaply, compute-wise</p><p>This property is called <strong>Hashing Efficiency</strong>. To solve it, we relied on hashing algorithms already available and proven in production use, ran extensive benchmarks, and adopted one as our hashing algorithm.</p><p>Should be able to run multiple non-overlapping experiments</p><p>With the Mutex layer sitting on top, we run non-overlapping experiments by allocating disjoint quotas to each experiment (20:30:50, etc.).</p><p><strong>What else?</strong></p><p><strong>SRM (Sample Ratio Mismatch)</strong> — since we rely heavily on hashing, we need guardrails around variant-distribution skew. Once the data becomes biased, weeks of data can be wasted because extra users landed in one variant.</p><p>So far, we’ve covered only the <strong>distribution</strong> side of experiments; we haven’t yet examined the <strong>analysis</strong> side, which determines whether observed metric uplifts are statistically significant. This is a large area that also addresses issues such as:</p><p><strong>Pre-experiment bias</strong> — Even when SRM and splits look fine, comparing metrics like orders (test) vs orders (control) can still reveal major bias. In distribution, each user counts equally: a user who trades once per week and a user who trades hundreds of times per day each count as one. From a metrics perspective, that causes serious problems unless we change our analysis methods.</p><p><strong>Coming Next</strong></p><p>How we solved SRM at AngelOne</p><p><strong>About the author:</strong></p><p>Rishabh Mishra</p><p>SDE 2 — Application Development</p><p><em>Angel One Limited; SEBI Regn. No: INZ000161534</em></p><p><em>Disclaimer: Investments in securities market are subject to market risks, read all the related documents carefully before investing </em><a href="https://bit.ly/4u4zuXN"><em>https://bit.ly/4u4zuXN</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0b10d3a367bd" width="1" height="1" alt=""><hr><p><a href="https://medium.com/angel-tech/how-angel-one-does-a-b-experimentations-at-scale-0b10d3a367bd">How Angel One does A/B Experimentations at Scale</a> was originally published in <a href="https://medium.com/angel-tech">Angel One Square</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Closing Price Just Got More Transparent. Here’s What That Means for You.]]></title>
            <link>https://medium.com/angel-tech/the-closing-price-just-got-more-transparent-heres-what-that-means-for-you-e766439d876f?source=rss-b06a776e75f4------2</link>
            <guid isPermaLink="false">https://medium.com/p/e766439d876f</guid>
            <category><![CDATA[product-management]]></category>
            <category><![CDATA[future-and-options]]></category>
            <category><![CDATA[fintech]]></category>
            <category><![CDATA[stock-market]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Angel One]]></dc:creator>
            <pubDate>Mon, 31 Aug 2026 02:43:38 GMT</pubDate>
            <atom:updated>2026-08-31T02:56:15.540Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Starting August 3, 2026, the way India’s most-traded stocks close has changed. Angel One has already built it in.</em></p><p>For years, a quiet vulnerability sat at the heart of every equity market close.</p><p>The official closing price of a stock, the number that goes into your portfolio valuation, mutual fund NAV, index calculation, and F&amp;O settlement, wasn’t set by the weight of the market. It was set by a 30-minute volume-weighted average. And anyone willing to place large orders in the final minutes of that window could move the needle.</p><p>Not dramatically. Not obviously. But enough to matter, for funds, for institutions, for anyone whose position settlement depended on where a stock closed.</p><p>SEBI fixed this on August 3, 2026. We built the fix into Angel One before it went live.</p><p><strong>A Single Pool. A Single Price. No Last-Minute Fluctuations</strong></p><p>The new mechanism is called the <strong>Closing Auction Session (CAS)</strong>.</p><p>Instead of computing a closing price from the last 30 minutes of trade, the exchange now runs a dedicated 20-minute auction, from 3:15 to 3:35 PM, for every stock that has Futures &amp; Options contracts. Think Reliance, TCS, HDFC Bank, Infosys, and all Nifty 50 and Sensex constituents*. Every eligible buy and sell order gets pooled together. The exchange finds the single price at which the most shares can actually change hands. That’s the close.</p><p>One pool. One price. Everyone transacts at it.</p><p>The result is a closing price that reflects genuine collective market consensus, not the momentum of a few large orders in the final seconds. The New York Stock Exchange, the London Stock Exchange, and Euronext have all operated this way for years. As of this month, so does India.</p><p>* The securities are quoted as an example and not as a recommendation</p><p><strong>Your Trading Day, Minute by Minute</strong></p><p>The mechanics are worth understanding, because the timing is specific.</p><p><strong>9:15–3:15 PM — Normal trading.</strong> The last 15 minutes of regular trading also serve a second function: they generate the <strong>Reference Price, </strong>the VWAP of all trades in this window. Think of it as the market’s last read before the auction begins. The exchange publishes it around 3:15 PM.</p><p><strong>3:15–3:20 PM — Transition window.</strong> The exchange closes for order entry while it computes the Reference Price and shifts into auction mode. (At Angelone, you can still place orders but they will be scheduled for 3:20 pm). One important detail: any Stop Loss or Iceberg orders, you had open are cancelled here, these order types aren’t permitted in the auction under SEBI rules. Unexecuted limit orders within allowed price range carry forward automatically and keep their time priority.</p><p><strong>3:20–3:30 PM — The order entry window</strong> This is the auction. You can place, modify, and cancel orders here. There’s a twist: the window doesn’t close at exactly 3:30 PM. It closes at a <strong>randomly chosen moment between 3:28 and 3:30 PM</strong>. This is deliberate, it prevents participants from flooding the book at the last second to influence the equilibrium price.</p><p>From 3:20 to 3:25 PM, both limit and market orders are accepted. From 3:25 PM until the window closes, only limit orders can be placed at the exchange level.</p><p><strong>3:30–3:35 PM — Matching</strong> No trades happen during the order entry window. Every order is held. At 3:30 PM, the exchange scans every price level within ±3% of the Reference Price, finds the price where the most shares can trade, and executes everything at once. That price becomes the official close. All matched orders — regardless of the price they were placed at, execute at this single price called equilibrium price.</p><p><strong>After 3:35 PM</strong> The closing price is published. Standard circuit limits are restored. And if you trade derivatives: the F&amp;O segment now runs until <strong>3:40 PM</strong>, giving you time to adjust positions once the CAS close is known.</p><p><strong>What Angel One Shows You During the Auction to help you take a decision to place orders</strong></p><p>During the auction, the usual landmarks change. Here’s what you’ll see on the app, and why each one is there.</p><p><strong>The Indicative Close Price replaces the LTP</strong></p><p>Once 3:15 PM hits, the Last Traded Price for CAS-eligible stocks stops updating. No new trades are happening in the cash segment. Displaying a stale LTP would be misleading especially at a time derivatives continue to trade.</p><p>Angel One replaces it with the <strong>Indicative Close Price</strong> — a live signal broadcast continuously by the exchange showing the price at which the maximum number of shares could be matched <em>right now</em>, given the current order book. It <strong>updates in real time using a web socket </strong>as orders arrive, are modified, or are cancelled. It’s the most accurate read of where the stock is heading into the close.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/240/1*vQ2q-SG3IIczLk2LK5fZhw.png" /><figcaption>* The securities are quoted as an example and not as a recommendation</figcaption></figure><p>This matters most for options traders. The official Nifty and Sensex values freeze at 3:15 PM while the auction runs, because their constituents aren’t generating new trade prices. Most platforms show a static index during this window. Angel One doesn’t. We compute and display a live <strong>Indicative Index Close Value</strong> from the auction data, so you have a real-time reference for pricing index options all the way through to 3:35 PM.</p><p><strong>Three signals that tell you how the auction is developing</strong></p><p>The stock detail page for any CAS-eligible stock shows three live data points throughout the order entry window:</p><p><strong>Indicative Matched Quantity</strong> — the volume that would execute at the current equilibrium price if the auction ended right now. High volume means strong two-sided participation. Low volume means the price may shift before matching concludes.</p><p><strong>Indicative Imbalance Quantity</strong> — the difference between cumulative buy and sell interest at the current equilibrium price, with a Buy-heavy or Sell-heavy tag. A strong buy-heavy imbalance tends to push the equilibrium price up as sellers are drawn in; sell-heavy pushes it down.</p><p><strong>Indicative Imbalance Quantity at Market</strong> — the same measure but restricted to market orders only. Market order participants are typically transacting with conviction, not on price sensitivity. This signal tells you where the firm directional money is sitting.</p><p>Together, these three numbers give you a continuous read on the auction’s direction, the likely volume that will clear, and the balance of informed flow, all updating in real time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/228/1*QVdbxW0K5f8kmCgXItQgvw.png" /><figcaption>* The securities are quoted as an example and not as a recommendation</figcaption></figure><p><strong>The CAS price band replaces circuit limits based on Reference Price</strong></p><p>From 3:15 PM, standard upper and lower circuit limits are replaced by a ±3% band around the Reference Price. Any order outside this range is rejected by the exchange. Angel One displays this band along with the reference price clearly on the order pad and stock detail page so you always know the valid range before placing an order.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/235/1*r-uR0CBAQPBRJfpSVqyDIQ.png" /><figcaption>* The securities are quoted as an example and not as a recommendation</figcaption></figure><p><strong>Two Ways Angel One Makes CAS Smoother Than the Exchange Rules Suggest</strong></p><p>The auction’s timing rules create two friction points. We’ve built around both of them.</p><p><strong>Schedule your order from 3:15 PM — don’t wait for 3:20</strong></p><p>The exchange won’t accept orders during the 3:15–3:20 PM transition window. Without Angel One’s handling, this means sitting and waiting, then racing to enter your order the instant the window opens.</p><p>Instead: place your order anytime from 3:15 PM. Angel One queues it, stamps it 3:20, and fires it to the exchange the moment the order entry window opens, automatically. Modify or cancel it freely until market close.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/721/1*HCz2QdzHTbk65tm8nwMEZw.png" /><figcaption>* The securities are quoted as an example and not as a recommendation</figcaption></figure><p>The practical upside: <strong>time priority</strong>. In CAS, all else equal, earlier orders are matched first. Getting your order to the exchange at the opening second of the window puts you ahead of anyone who waited. (Orders carried forward from regular trading rank even earlier, they retain their original pre-auction timestamps.)</p><p>It also eliminates the risk of missing the window entirely. Since the window closes at a random moment between 3:28 and 3:30 PM, scheduling early means you’re already in, regardless of when the cut-off falls.</p><p><strong>Place market orders through the entire auction window</strong></p><p>From 3:25 PM, the exchange only accepts limit orders. Market orders are locked out.</p><p>Angel One converts your market order into a <strong>protected limit order</strong> at a capped price before submitting it to the exchange. This means your market order reaches the exchange as a valid limit order, accepted even in the limit-only phase, while ensuring you can’t get a wildly adverse fill.</p><p>In practice: you can place, modify, and cancel market orders from 3:20 PM all the way through to 3:30 PM, without worrying about the exchange’s phase cutoffs. You get the simplicity of a market order, protection from an outlier execution, and no interrupted workflow.</p><p><strong>For Stocks without F&amp;O contracts, nothing changes</strong></p><p>Stocks without F&amp;O contracts are completely unaffected. They continue trading until 3:30 PM, and their closing price calculation remains exactly as it was.</p><p><strong>The Bottom Line</strong></p><p>The Closing Auction Session is a structural improvement to how India’s markets find the closing price. It removes the disproportionate influence of last-minute orders and replaces it with a single, exchange-wide price discovery process that every participant enters on equal footing.</p><p>For most users, the change is seamless, Angel One surfaces the right information at the right moment and handles the exchange’s timing mechanics in the background. The order entry workflow is familiar. The data on screen is richer.</p><p>The close is more transparent now. And on Angel One, it’s been live for several weeks.</p><p>About the Author:</p><p>Mohit Manaktala<br>Director of Product, Angel One</p><p><em>CAS applies to all F&amp;O-eligible stocks from August 3, 2026. For questions about specific order types or how your existing positions are affected, reach out to Angel One support.</em></p><p><em>Angel One Limited; SEBI Regn. No: INZ000161534</em></p><p><em>Disclaimer: Investments in securities market are subject to market risks, read all the related documents carefully before investing </em><a href="https://bit.ly/4u4zuXN"><em>https://bit.ly/4u4zuXN</em></a></p><p>For further reading, you can check the official Circulars and FAQs here:</p><p>SEBI Circular: <a href="https://www.sebi.gov.in/legal/circulars/jan-2026/introduction-of-closing-auction-session-cas-in-the-equity-cash-segment-and-certain-modifications-in-the-pre-open-auction-session_99122.html">https://www.sebi.gov.in/legal/circulars/jan-2026/introduction-of-closing-auction-session-cas-in-the-equity-cash-segment-and-certain-modifications-in-the-pre-open-auction-session_99122.html</a></p><p>NSE Circular: <a href="https://nsearchives.nseindia.com/content/circulars/FAOP74467.zip">https://nsearchives.nseindia.com/content/circulars/FAOP74467.zip</a></p><p>BSE Circular: <a href="https://www.bseindia.com/downloads/UploadDocs/Notices/20260728-42/20260728-42.pdf">https://www.bseindia.com/downloads/UploadDocs/Notices/20260728-42/20260728-42.pdf</a></p><p>FAQ: <a href="https://www.bseindia.com/downloads/UploadDocs/Notices/Attach/Annexurev1-FAQs$7526f61a-b66f-4a37-8bcb-cbdce64ac1ae.pdf">https://www.bseindia.com/downloads/UploadDocs/Notices/Attach/Annexurev1-FAQs$7526f61a-b66f-4a37-8bcb-cbdce64ac1ae.pdf</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e766439d876f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/angel-tech/the-closing-price-just-got-more-transparent-heres-what-that-means-for-you-e766439d876f">The Closing Price Just Got More Transparent. Here’s What That Means for You.</a> was originally published in <a href="https://medium.com/angel-tech">Angel One Square</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>