# JioHotstar

Product and Engineering notes from JioHotstar, India's leading OTT service

This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles.

## The Kafka Optimization That Saved Us $350K a Year

DevFeed: [The Kafka Optimization That Saved Us $350K a Year](<https://devfeed.tech/articles/the-kafka-optimization-that-saved-us-350k-a-year-89899.md>)

Original publisher: [Read original article](<https://blog.hotstar.com/the-kafka-optimization-that-saved-us-350k-a-year-48a0e05c3c08?source=rss----dbc3fcbc7f07---4>)

Author: Mayank

Published: 2026-10-19T07:37:31Z

Content type: article

Language: en

Sources: [JioHotstar](<https://devfeed.tech/sources/jiohotstar.md>)

Topics: [Apache Kafka](<https://devfeed.tech/topics/apache-kafka.md>), [data-engineering](<https://devfeed.tech/topics/data-engineering.md>), [Event-Streaming](<https://devfeed.tech/topics/event-streaming.md>), [Data serialization](<https://devfeed.tech/topics/data-serialization.md>), [zero downtime deployment](<https://devfeed.tech/topics/zero-downtime-deployment.md>)

Tags: [data](<https://devfeed.tech/tags/data.md>), [go](<https://devfeed.tech/tags/go.md>), [kafka](<https://devfeed.tech/tags/kafka.md>), [protobuf](<https://devfeed.tech/tags/protobuf.md>), [real-time](<https://devfeed.tech/tags/real-time.md>), [scale](<https://devfeed.tech/tags/scale.md>), [smt](<https://devfeed.tech/tags/smt.md>)

### AI overview

JioHotstar describes a stateless Kafka-to-Kafka service that scans protobuf wire bytes to retain only fields needed by downstream consumers. The article details offset tracking, delivery acknowledgments, backpressure, rebalancing, and shutdown behavior, reporting an average message-size reduction of 84.9% and close to $350,000 in annual savings.

### Source excerpt

The JioHotstar platform receives in excess of 40TB data every day. At 2.1 million messages per second. At this scale, every superflous byte is costly, like real money costly. In this blog, we cover strategies that helped us optimise our ingest workflow which ultimately saved real money. The Waste Our video streaming platform emits a telemetry heartbeat for every active playback session -- a rich, 7.5 KB protobuf event carrying 300+ fields of session context. Six independent downstream consumers subscribe to it. Most of them need about 1.1 KB. That is 6.4 KB of waste, multiplied by six consumers, multiplied by 2.1 million times a second. The annual cost ran into hundreds of thousands of dollars in Kafka egress alone -- before counting the CPU and GC pressure of fully deserializing messages just to throw away 85% of the fields. The Strategy We needed a way to strip the fat after ingestion but before fan-out, without losing a single message. The obvious tool -- Kafka Connect SMTs -- could not do it without reintroducing the full deserialization cost. So we built a stateless Kafka-to-Kafka micro service that scans raw protobuf wire bytes, extracts only the fields consumers actually need, and publishes a minimized event -- with a zero-loss offset tracking chain that guarantees at-least-once delivery through crashes and rebalances. This post walks through the architecture, the ack chain that makes zero loss possible, the wire-level protobuf trick that avoids proto.Unmarshal, and the rebalancing and shutdown sequences that keep it safe at scale. What Is a Heartbeat? A heartbeat is a periodic telemetry event emitted by every client on our video streaming platform during playback. It is the signal that tells our systems whether a session is healthy -- whether playback started, stalled, buffered, dropped quality, or failed. Our real-time alerting system, QoS dashboards, churn models, and several other systems all depend on it to understand the real-time health of every active sessi

## What We Learned from Benchmarking Knowledge Bases for Coding Agents

DevFeed: [What We Learned from Benchmarking Knowledge Bases for Coding Agents](<https://devfeed.tech/articles/what-we-learned-from-benchmarking-knowledge-bases-for-coding-agents-77971.md>)

Original publisher: [Read original article](<https://blog.hotstar.com/what-we-learned-from-benchmarking-knowledge-bases-for-coding-agents-6bbdd9e53727?source=rss----dbc3fcbc7f07---4>)

Author: Abhishek Bansal

Published: 2026-09-28T12:03:06Z

Content type: article

Language: en

Sources: [JioHotstar](<https://devfeed.tech/sources/jiohotstar.md>)

Topics: [agentic-coding](<https://devfeed.tech/topics/agentic-coding.md>), [Developer experience](<https://devfeed.tech/topics/developer-experience.md>), [Embeddings](<https://devfeed.tech/topics/embeddings.md>), [sourcegraph](<https://devfeed.tech/topics/sourcegraph.md>), [Tool](<https://devfeed.tech/topics/tool.md>)

Tags: [agentic-rag](<https://devfeed.tech/tags/agentic-rag.md>), [agents](<https://devfeed.tech/tags/agents.md>), [benchmark](<https://devfeed.tech/tags/benchmark.md>), [benchmarking](<https://devfeed.tech/tags/benchmarking.md>), [coding-agents](<https://devfeed.tech/tags/coding-agents.md>), [cost](<https://devfeed.tech/tags/cost.md>), [knowledge-base](<https://devfeed.tech/tags/knowledge-base.md>), [llm](<https://devfeed.tech/tags/llm.md>), [programming](<https://devfeed.tech/tags/programming.md>), [retrieval](<https://devfeed.tech/tags/retrieval.md>), [semantic](<https://devfeed.tech/tags/semantic.md>), [token-usage](<https://devfeed.tech/tags/token-usage.md>)

### AI overview

In JioHotstar's selected benchmark runs, retrieval helped according to task shape: semantic indexing performed best when prompts were vague and agents had to map intent to code, while filesystem navigation and structural search stayed competitive on concrete, well-scoped tasks. Sourcebot was more accurate and cheaper than the filesystem baseline on the hardest SWE-bench Verified tasks; on simpler Lite tasks, retrieval generally added cost and turns for similar results. Most retrieval modes did not save tokens by default, and indexing overhead was part of the evaluation. The team also found that agents favored familiar, composable tools such as grep, symbol lookup, and file reads. The evaluation used subsets run locally, so its findings do not establish a universal winner.

### Source excerpt

Introduction/TL;DR At JioHotstar, we have been thinking a lot about how to give developers a better AI-assisted engineering experience. Once coding agents started becoming useful for day-to-day work, the next natural question was: how do we help those agents understand our codebases better? That led us into the world of knowledge bases, repo indexers, code graphs, and semantic search tools. Almost every product in this space promises some version of the same outcome: better context, fewer blind searches, lower token usage, and a smoother developer experience. That sounded plausible, but we wanted to answer it with data instead of anecdotes. So we ran a practical evaluation across multiple frameworks, benchmarks, and retrieval styles. We were not trying to prove that one category of tool is always right. We were trying to answer a more useful engineering question: What kind of retrieval helps for what kind of problem, and what does it cost us in practice? The questions we cared about were straightforward: How useful are these knowledge bases and indexers for real programming tasks? Do they actually reduce token usage and cost in practice? We also wanted to evaluate them the way engineers really experience them, not just as a pure retrieval benchmark. So we looked at four things: Accuracy: did the agent actually solve the task? Token consumption / cost: what did it cost to get there? Agent turns: how long did the reasoning loop run? Agent friendliness: how naturally did the agent use the tools? This was also a practical engineering exercise, not a giant benchmark campaign running on a shared cluster. We used carefully chosen subsets, because full benchmark sweeps can run into hundreds of problems, and for knowledge-base-style systems the indexing overhead itself can become a serious part of the cost. The short version is: these systems can help, but their value depends heavily on the shape of the task. For well-defined engineering tasks, especially when the issue can

## Journey of an Ad Request: The Hidden Engineering

DevFeed: [Journey of an Ad Request: The Hidden Engineering](<https://devfeed.tech/articles/journey-of-an-ad-request-the-hidden-engineering-77967.md>)

Original publisher: [Read original article](<https://blog.hotstar.com/journey-of-an-ad-request-the-hidden-engineering-e7b7f9921e46?source=rss----dbc3fcbc7f07---4>)

Author: Ayush Kumar

Published: 2026-07-20T10:22:24Z

Content type: article

Language: en

Sources: [JioHotstar](<https://devfeed.tech/sources/jiohotstar.md>)

Topics: [streaming data analytics](<https://devfeed.tech/topics/streaming-data-analytics.md>), [Electrical engineering and computer science (EECS)](<https://devfeed.tech/topics/electrical-engineering-and-computer-science-eecs.md>)

Tags: [ads](<https://devfeed.tech/tags/ads.md>), [adtech](<https://devfeed.tech/tags/adtech.md>), [advertising](<https://devfeed.tech/tags/advertising.md>), [auction](<https://devfeed.tech/tags/auction.md>), [bidding](<https://devfeed.tech/tags/bidding.md>), [client](<https://devfeed.tech/tags/client.md>), [complexity](<https://devfeed.tech/tags/complexity.md>), [content](<https://devfeed.tech/tags/content.md>), [cpc](<https://devfeed.tech/tags/cpc.md>), [cue](<https://devfeed.tech/tags/cue.md>), [demand](<https://devfeed.tech/tags/demand.md>), [display](<https://devfeed.tech/tags/display.md>), [engine](<https://devfeed.tech/tags/engine.md>), [engineering](<https://devfeed.tech/tags/engineering.md>), [event](<https://devfeed.tech/tags/event.md>), [formats](<https://devfeed.tech/tags/formats.md>), [frontend](<https://devfeed.tech/tags/frontend.md>), [handshake](<https://devfeed.tech/tags/handshake.md>), [live](<https://devfeed.tech/tags/live.md>), [movie](<https://devfeed.tech/tags/movie.md>)

### AI overview

JioHotstar describes ad selection as a sub-100-millisecond process that filters candidates by user, device, content, location and advertiser constraints, checks frequency caps, paces delivery with PID or SHALE algorithms, then fills an ad pod through priority tiers. Direct deals return a VAST response immediately; programmatic ads depend on third-party SSP auctions and their uncertain fill rates. The article contrasts SSAI, which stitches ads into streams and may personalize by cohort, with CSAI, which uses advance cue markers, and SGAI, where the server guides player-side insertion for one-to-one personalization without advance markers. Tracking pipelines record delivery events, while click-through and view-through attribution credit conversions after clicks or ad exposure, respectively.

### Source excerpt

Advertising is a key source of revenue in most businesses. It certainly forms a key pillar for JioHotstar. There is so much technology hidden under the surface. In this blog, we will cover some of the complexity that gets unlocked when we render an ad on our content. Every 30-second ad break is a high-stakes performance involving millions of calculations, real-time bidding, and surgical precision. When you open JioHotstar to watch a live sports event or simply a movie, you see a Fence* ad (static display) or a Preroll* or a Midroll* Ad (video). What you don't see is the "Ad Decision" engine firing at millisecond speeds to decide: Why this ad? Why now? And why you? Speak Like an Adtech Pro Getting your foot inside the door, we need to understand some jargon! CPM (Cost Per Mille): The cost an advertiser pays for every 1,000 impressions. CPC -- Cost per Click SSAI (Server-Side Ad Insertion): Ads are "stitched" into the video stream on the server. At JioHotstar scale, this often happens at a Cohort level where we are grouping similar users to a balanced personalization of Ads. CSAI (Client Side Ad Insertion): The frontend client app(player) interrupts the video stream based on cue markers to show an ad. These cue markers are known well in advance when the user plays any content. SGAI (Server-Guided Ad Insertion): The server "guides" the player on what to do, but the player performs the final insertion. This allows for 1:1 personalization at a live event scale where we do not have the cue markers in advance. Programmatic vs. Direct: Direct is a handshake deal (e.g., "I want 1M views on an Ad"). Programmatic is an automated auction where we ask a 3rd-party SSP (Supply Side Platform) for an ad in real-time. Some common Ad Formats on JioHotstar Preroll Ad: The video ad you see before the actual content video starts. Midroll Ad: The video ad you see while in between the content. Fence Ad: The static display ad you see below the player on the app. The Demand Puzzle Imagine you

## Consistent Generative Video: Building Video Models That Remember

DevFeed: [Consistent Generative Video: Building Video Models That Remember](<https://devfeed.tech/articles/consistent-generative-video-building-video-models-that-remember-77966.md>)

Original publisher: [Read original article](<https://blog.hotstar.com/getting-past-the-first-frame-building-video-models-that-remember-31f03d0710e1?source=rss----dbc3fcbc7f07---4>)

Author: Yaswanth Karnati

Published: 2026-07-01T06:53:14Z

Content type: article

Language: en

Sources: [JioHotstar](<https://devfeed.tech/sources/jiohotstar.md>)

Topics: [lora](<https://devfeed.tech/topics/lora.md>), [Fine-tuning](<https://devfeed.tech/topics/fine-tuning.md>), [Generative AI](<https://devfeed.tech/topics/generative-ai.md>), [Advanced AI workflows](<https://devfeed.tech/topics/advanced-ai-workflows.md>), [multimodal-ai](<https://devfeed.tech/topics/multimodal-ai.md>), [data-engineering](<https://devfeed.tech/topics/data-engineering.md>)

Tags: [0-1](<https://devfeed.tech/tags/0-1.md>), [audio](<https://devfeed.tech/tags/audio.md>), [challenges](<https://devfeed.tech/tags/challenges.md>), [consistency](<https://devfeed.tech/tags/consistency.md>), [data](<https://devfeed.tech/tags/data.md>), [diffusion](<https://devfeed.tech/tags/diffusion.md>), [fine-tuning](<https://devfeed.tech/tags/fine-tuning.md>), [generative-ai](<https://devfeed.tech/tags/generative-ai.md>), [generative-video](<https://devfeed.tech/tags/generative-video.md>), [identity](<https://devfeed.tech/tags/identity.md>), [ip](<https://devfeed.tech/tags/ip.md>), [lora](<https://devfeed.tech/tags/lora.md>), [ltx](<https://devfeed.tech/tags/ltx.md>), [models](<https://devfeed.tech/tags/models.md>), [pipeline](<https://devfeed.tech/tags/pipeline.md>), [production](<https://devfeed.tech/tags/production.md>), [video](<https://devfeed.tech/tags/video.md>), [wan-2-1](<https://devfeed.tech/tags/wan-2-1.md>)

### AI overview

The excerpt describes a LoRA pipeline for generating talking-head videos of catalog characters from a reference image and audio. LoRA freezes base weights and adds small character adapters (about 400 MB at rank 32, versus a 41 GB model), which can be swapped without reloading the base. Training data is divided into scenes, filtered for the target face, balanced by shot type, captioned with a character trigger, and precomputed into latents and text embeddings; face retrieval achieved about 85% recall. Wan 2.1's single-stage adapter produced robotic motion and identity drift; three-stage training and inference tuning improved results but left lip-sync and motion below production needs. LTX-2 worked with a simpler single-stage adapter and showed better identity retention, lip-sync, motion, and generalization to unseen scenes. Its best checkpoints typically fell at 6k-9k steps, while later training risked memorizing poses. Veo 3.1 and Kling 3.0 were compared without LoRA because they do not support adapter injection, limiting their use for reusable character-specific generation.

### Source excerpt

Introduction Generative AI offers a path to content creation at speeds and scales that traditional pipelines cannot match. One high-value use case: making existing IP characters-sports presenters, show hosts, brand ambassadors -deliver new scripts in their own likeness, without a reshoot, within IP & contract limitations. The premise is simple: take a reference image, provide audio, and generate a video where the character speaks naturally. In practice, this is harder than it sounds. Base proprietary video models like Kling, Veo often struggle with lip-sync quality, identity drift, and motion consistency. To address this, we built a production pipeline for character-consistent talking-head generation using LoRA (Low-Rank Adaptation), a lightweight fine-tuning method that adapts a base model to a specific character without retraining the entire network. This makes character adaptation practical to train, store, and deploy across multiple IP characters at scale. Why LoRA? LoRA works by freezing the original model weights and injecting small, trainable low-rank matrices into specific layers. Instead of updating a weight matrix W directly, LoRA decomposes the update into two smaller matrices: ΔW = A x B, where A and B have a rank r far smaller than the original dimensions. For rank 32 (our configuration), each adapted layer gains only ~0.1% additional parameters-producing a ~400MB adapter file vs. 41GB for the full model. Multiple character adapters can be swapped at inference without reloading the base model. LoRA is well-established for image generation, but video introduces harder challenges: the adapter must preserve identity across dozens of frames (not just one), motion and structure are entangled across diffusion noise levels, and video LoRAs train on far fewer samples (15-20 clips vs. hundreds of images), making overfitting the dominant risk. This blog covers how we fine-tuned two video architectures-Wan 2.1 and LTX-2-with LoRA, the data preparation that made it

## Beyond Imitation: Teaching AI to Speak Like a Character in Hindi

DevFeed: [Beyond Imitation: Teaching AI to Speak Like a Character in Hindi](<https://devfeed.tech/articles/beyond-imitation-teaching-ai-to-speak-like-a-character-in-hindi-77963.md>)

Original publisher: [Read original article](<https://blog.hotstar.com/beyond-imitation-teaching-ai-to-speak-like-a-character-in-hindi-476bd33f049a?source=rss----dbc3fcbc7f07---4>)

Author: Sagar Tekwani

Published: 2026-05-08T13:03:55Z

Content type: article

Language: en

Sources: [JioHotstar](<https://devfeed.tech/sources/jiohotstar.md>)

Topics: [Generative AI](<https://devfeed.tech/topics/generative-ai.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [generative-ai-tools](<https://devfeed.tech/topics/generative-ai-tools.md>), [Artificial Intelligence](<https://devfeed.tech/topics/ai.md>), [Uncertainty quantification LLMs](<https://devfeed.tech/topics/uncertainty-quantification-llms.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [architectures](<https://devfeed.tech/tags/architectures.md>), [chatterbox](<https://devfeed.tech/tags/chatterbox.md>), [generation](<https://devfeed.tech/tags/generation.md>), [generative-ai-use-cases](<https://devfeed.tech/tags/generative-ai-use-cases.md>), [indic](<https://devfeed.tech/tags/indic.md>), [reliability](<https://devfeed.tech/tags/reliability.md>), [scale](<https://devfeed.tech/tags/scale.md>), [speech-models](<https://devfeed.tech/tags/speech-models.md>)

### AI overview

The article describes JioHotstar's work on character-consistent Hindi and Hinglish speech generation. It details emotion-focused data preparation, fine-tuning the T3 text-to-token model while freezing other components, and mitigations for short-phrase artifacts and code-switched pronunciation errors. The reported results include improved speaker similarity, emotion alignment and pronunciation accuracy, with MOS rising from 3.3 to 4.2.

### Source excerpt

Introduction As JioHotstar serves a vast, diverse audience, generative AI is crucial for fast, scalable content creation across languages and formats, capabilities that traditional methods cannot match. Content systems are rapidly shifting to generative-first architectures, synthesising video, audio, and narrative programmatically. However, progress is uneven: visual generation is advancing rapidly, while audio generation lags, especially for Indic languages, which lack the mature datasets and investment seen in English speech. Under realistic production workloads in cinematic contexts, existing speech systems struggle to preserve emotional expressiveness, maintain pronunciation stability, and deliver consistent outputs across repeated generations. These failures persist even on widely adopted commercial platforms, making audio the primary bottleneck in deploying AI-driven content pipelines alongside increasingly capable video systems. Why Indic Audio Breaks at Scale Speech generation systems that appear stable in demos often degrade under production-scale workloads. For Indic languages, these failures surface earlier and more acutely due to four structural constraints: Lack of emotion-rich training data: Indic speech datasets are dominated by neutral narration. Unlike English, which benefits from large emotionally diverse corpora and mature conditioning techniques, Indic models struggle to produce emotionally aligned speech in character-driven settings. Insufficiency of prompt-based control: Prompting influences tone superficially but cannot enforce consistency or repeatability. Without explicit modeling of emotion, language-specific phonetics, and stability constraints, prompt-level approaches fail production reliability requirements. Linguistic complexity and code-switching: Hindi speech relies heavily on prosody, vowel duration, and rhythmic stress. Most systems lack language-aware mechanisms to resolve these variations, producing inconsistent or unnatural Hingl

## Beyond Manual Checks: How We Built a 24/7 "Watchdog" for Live Sports Streaming

DevFeed: [Beyond Manual Checks: How We Built a 24/7 "Watchdog" for Live Sports Streaming](<https://devfeed.tech/articles/beyond-manual-checks-how-we-built-a-24-7-watchdog-for-live-sports-streaming-77964.md>)

Original publisher: [Read original article](<https://blog.hotstar.com/beyond-manual-checks-how-we-built-a-24-7-watchdog-for-live-sports-streaming-026fdcaebc5e?source=rss----dbc3fcbc7f07---4>)

Author: Abhishek Sinha

Published: 2026-04-08T12:43:23Z

Content type: article

Language: en

Sources: [JioHotstar](<https://devfeed.tech/sources/jiohotstar.md>)

Topics: [Test automation](<https://devfeed.tech/topics/test-automation.md>), [synthetic monitoring](<https://devfeed.tech/topics/synthetic-monitoring.md>), [Streaming](<https://devfeed.tech/topics/streaming.md>), [Production Engineering](<https://devfeed.tech/topics/production-engineering.md>), [health check](<https://devfeed.tech/topics/health-check.md>), [workflow automation](<https://devfeed.tech/topics/workflow-automation.md>)

Tags: [apis](<https://devfeed.tech/tags/apis.md>), [automated](<https://devfeed.tech/tags/automated.md>), [automation](<https://devfeed.tech/tags/automation.md>), [camera](<https://devfeed.tech/tags/camera.md>), [cloud-costs](<https://devfeed.tech/tags/cloud-costs.md>), [concurrent](<https://devfeed.tech/tags/concurrent.md>), [device](<https://devfeed.tech/tags/device.md>), [devices](<https://devfeed.tech/tags/devices.md>), [drm](<https://devfeed.tech/tags/drm.md>), [engineer](<https://devfeed.tech/tags/engineer.md>), [failed](<https://devfeed.tech/tags/failed.md>), [failure](<https://devfeed.tech/tags/failure.md>), [gap](<https://devfeed.tech/tags/gap.md>), [languages](<https://devfeed.tech/tags/languages.md>), [live](<https://devfeed.tech/tags/live.md>), [live-event](<https://devfeed.tech/tags/live-event.md>), [live-streaming](<https://devfeed.tech/tags/live-streaming.md>), [logic](<https://devfeed.tech/tags/logic.md>), [manifest](<https://devfeed.tech/tags/manifest.md>), [sports](<https://devfeed.tech/tags/sports.md>), [streaming](<https://devfeed.tech/tags/streaming.md>), [video-validation](<https://devfeed.tech/tags/video-validation.md>)

### AI overview

JioHotstar describes a live-stream validation system that follows events from scheduled matches through unpublishing. A metadata-driven orchestrator discovers matches, regional feeds and camera angles, starts workers before play, and stops them when content is unpublished; it validates stream variants independently and dispatches two concurrent checks. Video Validator inspects manifests and packets for discontinuities, missing segments, sequence drift, SSAI cue problems and inconsistent CDN chunks. Play Portal tests actual playback on Android, iPhone, TV and web devices, including startup, audio-language options and rebuffering. The article says Kubernetes autoscaling handles demand surges, while continuous checks can catch issues such as regional manifest 404s, ad cue failures and bitrate sync problems before publication or user impact. It contrasts server-side stream inspection with real-device playback, describing them as complementary because a healthy 200 OK manifest alone does not ensure a player works.

### Source excerpt

In the world of live sports streaming, there is no "undo" button. Live events bring out emotions in viewers and any minor buffering or failure can cause disappointment in our customers, especially if it happens in crunch moments of the live event. Our Live video validation has gradually moved from being fully manual to islands of automation. We had the tools to check manifests and APIs, but we lacked a cohesive brain to run them. Here we talk about how we moved from manual intervention to an automated, cohesive, real-device validation ecosystem. Manual "QA" Failed at Scale Before we built our current pipeline, our validation process was a series of stretch efforts that simply couldn't keep up with the scale of modern streaming. Manual QA can work for a single video and small number of matches, but when we look at streaming a live event in 12 languages, 5 camera angles, 74 matches over 2 months, automation is the only way to reliably run a QA process. At peak, JioHotstar (JHS) has seen close to 55 concurrent live streams, each with their unique stream variations. The Orchestration Gap: We started automated stream checks early on, but stream checks were started manually. If an engineer forgot to trigger a script for a specific regional feed, that feed went unmonitored. The "Orphaned" Test Problem: We would often find validation scripts running for hours after a match ended because there was no central logic to tell them to stop. Automation is wonderful but cloud costs are real too. The Silent Failure: A manifest might look perfect on a server (200 OK), but the actual player on a Smart TV might be stuck on a buffering wheel due a media sequence mismatch or due to ads related discontinuity tags or a DRM mismatch. We didn't just need a bigger QA team; we needed an orchestrator for our automation, a deeper video validation mechanism and real devices. Phase 1: The Steering (Video Validation Orchestrator) It's not enough for a manual or automated test to certify that a live

## Orchestrating JioHotstar Traffic: The Difference Between a Loading Spinner and a Winning Six

DevFeed: [Orchestrating JioHotstar Traffic: The Difference Between a Loading Spinner and a Winning Six](<https://devfeed.tech/articles/orchestrating-jiohotstar-traffic-the-difference-between-a-loading-spinner-and-a-winning-six-77970.md>)

Original publisher: [Read original article](<https://blog.hotstar.com/orchestrating-jiohotstar-traffic-the-difference-between-a-loading-spinner-and-a-winning-six-8a385b01380e?source=rss----dbc3fcbc7f07---4>)

Author: Karan Kaul

Published: 2026-03-09T09:41:53Z

Content type: article

Language: en

Sources: [JioHotstar](<https://devfeed.tech/sources/jiohotstar.md>)

Topics: [Load Balancing](<https://devfeed.tech/topics/load-balancing.md>), [network-intelligence](<https://devfeed.tech/topics/network-intelligence.md>), [observability](<https://devfeed.tech/topics/observability.md>), [Internet Traffic](<https://devfeed.tech/topics/internet-traffic.md>), [streaming data analytics](<https://devfeed.tech/topics/streaming-data-analytics.md>), [traffic splitting](<https://devfeed.tech/topics/traffic-splitting.md>)

Tags: [4g](<https://devfeed.tech/tags/4g.md>), [4k](<https://devfeed.tech/tags/4k.md>), [5g](<https://devfeed.tech/tags/5g.md>), [app](<https://devfeed.tech/tags/app.md>), [audacious](<https://devfeed.tech/tags/audacious.md>), [bandwidth](<https://devfeed.tech/tags/bandwidth.md>), [cdn](<https://devfeed.tech/tags/cdn.md>), [challenges](<https://devfeed.tech/tags/challenges.md>), [engineering](<https://devfeed.tech/tags/engineering.md>), [event](<https://devfeed.tech/tags/event.md>), [experience](<https://devfeed.tech/tags/experience.md>), [fiber](<https://devfeed.tech/tags/fiber.md>), [india](<https://devfeed.tech/tags/india.md>), [isp](<https://devfeed.tech/tags/isp.md>), [last-mile-delivery](<https://devfeed.tech/tags/last-mile-delivery.md>), [live](<https://devfeed.tech/tags/live.md>), [live-event](<https://devfeed.tech/tags/live-event.md>), [live-streaming](<https://devfeed.tech/tags/live-streaming.md>), [loading](<https://devfeed.tech/tags/loading.md>), [local](<https://devfeed.tech/tags/local.md>), [manage](<https://devfeed.tech/tags/manage.md>), [qos](<https://devfeed.tech/tags/qos.md>), [technology](<https://devfeed.tech/tags/technology.md>), [traffic](<https://devfeed.tech/tags/traffic.md>)

### AI overview

Hotstar's QoS Routing Manager steers streaming traffic to CDNs using performance data for cohorts defined by ASN, country, state, city, and user type. It uses ARGUS telemetry on playback failures, rebuffering, and RTT latency, smooths noisy measurements with an EWMA filter (α ~= 0.6), and weights playback reachability most heavily when scoring CDN health. Capacity rules throttle CDNs approaching limits and, at a utilization threshold, aim to exhaust CDN capacity uniformly. Safety gates enforce network availability constraints, cap per-iteration traffic shifts, and keep functional CDNs above a configurable minimum weight, typically 5%. In an A/B rollout using ASN-State cohorts during a T20 match, the treatment group saw an 11% improvement in playback failure rate and 2% improvements in rebuffering and RTT latency. The account describes this rollout and its results; it does not establish outcomes for other events or cohort configurations.

### Source excerpt

If you've ever streamed a high-stakes match on JioHotstar, the experience probably felt like a simple "tap and play". It's a seamless transition from the app icon to the stadium; a flick of a finger and you're right there, cheering for your players to hit another six. Beneath that simple play button lies one of the most aggressive engineering challenges in the world. At Hotstar, we navigate a "Tsunami" of traffic across one of the most complex network landscapes in the world. During events like the IPL or a high-stakes ODI match, we don't just manage millions of users - we manage millions of unique network realities. If you thought just placing a Content Delivery Network (CDN) makes the magic happen -- read on. The Illusion of the Monolithic Network Traditionally, CDN traffic management happens at the network provider or state level. For years, this was enough. But at our scale, "good enough" is the enemy of the audacious. A live event requires upwards of 60-80 Tbps of network bandwidth for streaming. To put things into perspective, it's like instantly grabbing the entire 4K movie collection from JioHotstar (hundreds of blockbusters). Now imagine finishing all those downloads in just one second ....and then doing it again the next second and the next. That is the relentless pace of the tidal wave we face during a major event. However, India's network isn't a monolith. It is a mosaic of fiber, 4G, 5G and fluctuating bandwidth that changes from one street to the next. A user on a 5G connection in South Delhi faces a vastly different network topology than a user on a local ISP in rural Rajasthan. When you are only as good as the video you deliver, you realize that macro-level routing hits a ceiling. To provide the best Quality of Service (QoS), we needed to treat every state, every city, and eventually every cohort, as a unique routing decision. The Solution: The QoS Routing Manager To solve the "Last Mile" problem, we built a real-time observability and orchestration eng

## Monetisation : Multi-period DASH x ExoPlayer

DevFeed: [Monetisation : Multi-period DASH x ExoPlayer](<https://devfeed.tech/articles/monetisation-multi-period-dash-x-exoplayer-77969.md>)

Original publisher: [Read original article](<https://blog.hotstar.com/monetisation-multi-period-dash-x-exoplayer-db8e8c00e521?source=rss----dbc3fcbc7f07---4>)

Author: Abhishek Bansal

Published: 2026-03-04T04:18:06Z

Content type: article

Language: en

Sources: [JioHotstar](<https://devfeed.tech/sources/jiohotstar.md>)

Topics: [ExoPlayer](<https://devfeed.tech/topics/exoplayer.md>), [Software Testing](<https://devfeed.tech/topics/software-testing.md>)

Tags: [ads](<https://devfeed.tech/tags/ads.md>), [alternatives](<https://devfeed.tech/tags/alternatives.md>), [android](<https://devfeed.tech/tags/android.md>), [approach](<https://devfeed.tech/tags/approach.md>), [asset](<https://devfeed.tech/tags/asset.md>), [audio](<https://devfeed.tech/tags/audio.md>), [box](<https://devfeed.tech/tags/box.md>), [card](<https://devfeed.tech/tags/card.md>), [client](<https://devfeed.tech/tags/client.md>), [complexity](<https://devfeed.tech/tags/complexity.md>), [component](<https://devfeed.tech/tags/component.md>), [consider](<https://devfeed.tech/tags/consider.md>), [content](<https://devfeed.tech/tags/content.md>), [cost](<https://devfeed.tech/tags/cost.md>), [dash](<https://devfeed.tech/tags/dash.md>), [details](<https://devfeed.tech/tags/details.md>), [driving](<https://devfeed.tech/tags/driving.md>), [dub](<https://devfeed.tech/tags/dub.md>), [exoplayer](<https://devfeed.tech/tags/exoplayer.md>), [period](<https://devfeed.tech/tags/period.md>), [streaming](<https://devfeed.tech/tags/streaming.md>)

### AI overview

ExoPlayer's AdsMediaSource rejected multi-period DASH streams, and simply removing its single-period assertion was insufficient: shared ad state could repeat prerolls across periods and misalign cue points. The team built a MultiPeriodAdTimeline that keeps one logical ad state, transforms cue-point times relative to each period, marks out-of-period groups skipped while preserving their indices, and uses bookmarking to prevent watched ads from replaying after cross-period seeks. This supports ads across period boundaries while allowing localized disclaimers and dub cards without re-encoding the main video. Older app versions continued receiving single-period DASH; multi-period delivery was gated to newer versions and rolled out with playback and ad metrics monitored. Separate players, playlist stitching, and concatenated clipped media sources were rejected because they add buffering, lose AdsMediaSource capabilities, and spread complexity.

### Source excerpt

Client Side Ad Insertion(CSAI) levels up using multi-period DASH in the Android ecosystem. This is our journey to upgrade ExoPlayer to leverage multi-period DASH.Introduction Multi-period DASH is a variant of DASH format that offers significant advantages, such as the ability to insert dynamic content like disclaimers, dub cards, and ad breaks without re-encoding the entire video. This flexibility is crucial for delivering personalized and localized content to millions of users across diverse regions. While the DASH standard natively supports multi-period manifests, the Android ExoPlayer ecosystem (specifically the AdsMediaSource component) was designed with a single-period assumption. This missing piece in the puzzle meant we couldn't support Client-Side Ad Insertion (CSAI) on these modern streams out of the box. This post details our engineering journey: identifying the constraints, evaluating architectural alternatives, and ultimately redesigning ExoPlayer's Ad handling to support multi-period content seamlessly. ContextThe Evolution of DASH Content JioHotstar relied on single-period DASH for entertainment content, where the entire stream including ads was encapsulated within a single period. This approach limited flexibility: inserting disclaimers, dub cards, and to some extent, dynamic ad breaks required re-encoding the entire video, which was both time-consuming and resource-intensive. Consider a common localization scenario: the same title is launched across multiple regions with different primary languages. Audio dubs may be recorded by different artists per market, so dub-card credits vary by region; likewise, legal and regulatory requirements differ, driving region-specific disclaimers. Under a single-period workflow, each of these variations would necessitate a distinct transcode of the full asset. If we target ~20 regions, that implies ~20 full transcodes, leading to long time-to-market, high cost, and significant operational complexity -- super unscalabl

## Gen AI Video -- Building Scalable Validation Framework

DevFeed: [Gen AI Video -- Building Scalable Validation Framework](<https://devfeed.tech/articles/gen-ai-video-building-scalable-validation-framework-77965.md>)

Original publisher: [Read original article](<https://blog.hotstar.com/building-scalable-validation-framework-for-video-generation-6c67d1177ce2?source=rss----dbc3fcbc7f07---4>)

Author: Sagar Tekwani

Published: 2026-02-18T03:42:23Z

Content type: article

Language: en

Sources: [JioHotstar](<https://devfeed.tech/sources/jiohotstar.md>)

Topics: [Generative AI](<https://devfeed.tech/topics/generative-ai.md>), [Video Analytics](<https://devfeed.tech/topics/video-analytics.md>), [human review](<https://devfeed.tech/topics/human-review.md>), [generative-ai-tools](<https://devfeed.tech/topics/generative-ai-tools.md>), [Image](<https://devfeed.tech/topics/image.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-validation](<https://devfeed.tech/tags/ai-validation.md>), [ai-video](<https://devfeed.tech/tags/ai-video.md>), [ai-video-generation](<https://devfeed.tech/tags/ai-video-generation.md>), [consistency](<https://devfeed.tech/tags/consistency.md>), [generative-ai-use-cases](<https://devfeed.tech/tags/generative-ai-use-cases.md>), [generative-video](<https://devfeed.tech/tags/generative-video.md>), [pipeline](<https://devfeed.tech/tags/pipeline.md>), [quality](<https://devfeed.tech/tags/quality.md>), [safety](<https://devfeed.tech/tags/safety.md>), [validation](<https://devfeed.tech/tags/validation.md>), [video](<https://devfeed.tech/tags/video.md>)

### AI overview

JioHotstar describes a validation framework embedded throughout its generative video pipeline. It checks intermediate outputs, sends go/no-go signals, and triggers regeneration or parameter changes; brand and safety violations are hard gates, with repeated failures escalated for manual review. Character identity is checked against creative-approved reference images using an ensemble of facial models and cosine similarity, with thresholds calibrated and periodically updated from human feedback. For deformities, the team reports that Gemini and Qwen-VL reached about 40% recall on its annotated dataset; a curated-data YOLO-family detector improved recall by about 35% relative to those models. Location checks use required-object detection and text-image alignment, plus validated keyframes as anchors for recurring locations. Brand and unsafe-content detection reports over 95% recall on internal evaluations. The authors say long-range story continuity remains experimental, and subtle spatial inconsistencies are not yet reliably detected.

### Source excerpt

Gen AI Video -- Building A Scalable Validation FrameworkWe discuss building strong eval frameworks as part of our Generative AI studio to ensure that generated video maintains a high quality bar.Introduction As JioHotstar scales to serve one of the world's largest streaming audiences, generative AI(GenAI) offers a powerful opportunity to accelerate content creation, adapt stories across languages and regions, and unlock creative workflows that traditional pipelines cannot match. Customers evaluate AI-generated video with the same standards they apply to premium productions; any drift in character identity, structural deformity, abrupt scene shift, or unsafe visual undermines realism instantly. Generative models, being probabilistic, introduce such inconsistencies naturally, and at our scale even rare defects accumulate into meaningful quality gaps. Ensuring stable characters, coherent locations, and safe content therefore becomes a scientific challenge central to customer acceptance. To address this, we built a Validation Framework that operates as a first-class component of the generation pipeline. This closed-loop system enforces quality, safety, and consistency at the same cadence as creation, enabling generative video to meet production-grade expectations on the JioHotstar platform. System Overview: The Validation Layer The video generation process starts with editorial scripts, from which the system extracts characters, locations, accessories, and context. These entities drive keyframe generation, which expands into video clips and assembles into the final video. The Validation Framework (Fig. 1) spans all stages of this process and operates synchronously within the workflow. It evaluates intermediate outputs, flags issues early, and triggers targeted regeneration or parameter adjustments to maintain quality. Brand detection and safety checks act as hard gates, while other modules guide regeneration to preserve consistency and visual integrity. Each module gover

## Modernizing Dependency Management: Beyond CocoaPods (Part 2 -- The Execution)

DevFeed: [Modernizing Dependency Management: Beyond CocoaPods (Part 2 -- The Execution)](<https://devfeed.tech/articles/modernizing-dependency-management-beyond-cocoapods-part-2-the-execution-77968.md>)

Original publisher: [Read original article](<https://blog.hotstar.com/modernizing-dependency-management-beyond-cocoapods-part-2-the-execution-bf3b9d739efb?source=rss----dbc3fcbc7f07---4>)

Author: Saurabh Kapoor

Published: 2026-02-13T04:26:22Z

Content type: article

Language: en

Sources: [JioHotstar](<https://devfeed.tech/sources/jiohotstar.md>)

Topics: [Dependency management](<https://devfeed.tech/topics/dependency-management.md>), [cocoapods](<https://devfeed.tech/topics/cocoapods.md>), [swift-package](<https://devfeed.tech/topics/swift-package.md>), [CI/CD](<https://devfeed.tech/topics/cicd.md>)

Tags: [cocoapods](<https://devfeed.tech/tags/cocoapods.md>), [dependency-management](<https://devfeed.tech/tags/dependency-management.md>), [ios-development](<https://devfeed.tech/tags/ios-development.md>), [migration](<https://devfeed.tech/tags/migration.md>), [mobile-development](<https://devfeed.tech/tags/mobile-development.md>), [spm](<https://devfeed.tech/tags/spm.md>), [swift](<https://devfeed.tech/tags/swift.md>), [swift-package-manager](<https://devfeed.tech/tags/swift-package-manager.md>)

### AI overview

The team migrated 60 CocoaPods to Swift Package Manager over six months in four phases, keeping CocoaPods available until releases succeeded without it. They first tested coexistence with mature third-party packages while preserving developer workflows and CI. For private binary targets, SPM's downloads lacked authentication; configuring `.netrc` credentials on developer machines and CI enabled authenticated downloads and checksum verification. Development pods became packages with preserved module boundaries, separate Objective-C and Swift targets, explicit resources and build settings, and most platform-specific XIB UI moved into code. CocoaPods `post_install` logic shifted to scoped Xcode build phases or explicit configuration. The migration also exposed runtime duplicate-symbol and framework-loading issues, plus Slather's lack of first-class support for pure Swift packages; the team audited linkage and moved coverage to an `xcresult`-based pipeline. The account emphasizes phased validation and the tradeoff of upfront refactoring for more explicit package and build behavior.

### Source excerpt

Modernizing Dependency Management: Beyond CocoaPods (Part 2 -- The Execution)Recap In Part 1, we shared the strategy -- why we had to move away from CocoaPods, how we evaluated our options, and the phased migration plan we designed to modernize 60 pods over 2 quarters without disrupting feature development. We bring it all home and share our learnings and scars. Phase 1: Proving That SPM Could Coexist Big bang approaches for a product that supports over a billion Indians is not on the menu. We could not stop the train and we had to upgrade while we ran as fast as we were. Swift Package Manager (SPM) had to live alongside CocoaPods without disrupting day-to-day development. This was not negotiable. It was also about building confidence gradually. Our dependency graph was already complex, and CocoaPods was deeply embedded in how the app was built, tested, and shipped. Introducing a second dependency manager into that ecosystem was risky. If coexistence failed, everything downstream would be compromised. So we treated Phase 1 as a controlled experiment. Defining Success We were explicit about what success looked like: Developers should not need to change their workflows CI pipelines must remain stable No runtime regressions CocoaPods must remain fully functional Choosing the Right First Dependencies We intentionally avoided internal SDKs in this phase. Instead, we chose third-party libraries that already had mature SPM support and met three criteria: widely used in the app, minimal customization, and no deep runtime coupling with internal frameworks or development pods. This allowed us to isolate SPM behavior without risking business-critical flows. By migrating only a handful of carefully selected dependencies, we could observe real-world behavior without destabilizing the system. Invisible to Developers One of the most important constraints we enforced was developer invisibility. Developers continued to open the same workspace, build using the same schemes, run tests t