SShortSingh.
Back to feed

Go and FFmpeg Service Cuts Mobile LCP from 4.1s by Owning Video Thumbnail Generation

0
·1 views

A development team behind TrendVidStream rebuilt their video thumbnail pipeline after identifying that hotlinked third-party 1280x720 JPEGs were wasting roughly 90KB of transfer per card across 24 cards and eight regional page variants. Their PHP 8.4 shared-hosting front end was unsuitable for thumbnail extraction due to execution time caps, disabled shell access, and no concurrency control between cron jobs. To solve this, they built a lightweight Go service that uses FFmpeg to extract representative frames from source videos and re-encodes them at three widths in WebP format, writing outputs to content-addressed paths. Go was chosen specifically because it supports a long-lived bounded worker pool, reliable context deadlines for killing stalled FFmpeg processes, and in-process deduplication to prevent redundant decodes. The team also detailed critical FFmpeg flags — including input-side seeking, -nostdin, single-frame extraction, and the thumbnail filter — as essential to making the service production-safe and performant.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingHacker News ·

Anthropic's Claude AI Models Experience Degraded Performance Outage

Anthropic reported a service disruption affecting multiple Claude AI models, as noted on the platform's official status page. The incident was logged and made publicly visible through Claude's status tracking system. The issue drew attention from the tech community, appearing briefly on Hacker News with modest engagement. No detailed cause or timeline for resolution was provided in the available information. Anthropic's status page remained the primary source for updates on the ongoing degradation.

0
ProgrammingHacker News ·

machine0 Launches Persistent CLI-Based Cloud VMs for AI Agent Workloads

Y Combinator startup machine0, founded by Barnaby, has launched a command-line tool that provisions persistent cloud virtual machines designed for long-running AI agent tasks. The service offers VMs billed by the minute, starting at $0.013 per hour, scaling up to 60 vCPUs, 240 GB RAM, and high-end GPUs including H100s and H200s. Each VM is a full KVM instance with 99.99% uptime, supporting features like snapshots, block storage, and credential profiles to isolate agent environments. The platform is already used by teams running agent fleets, reinforcement learning workloads, and product infrastructure where thousands of short-lived machines are spun up from versioned images. machine0 aims to address key challenges of running AI agents at scale, including resource constraints, security isolation, and task continuity when local machines go offline.

0
ProgrammingDEV Community ·

Google Tests AI Mode Shortcuts for Images, Files and Brainstorming in Desktop Search

Google is running a limited test of three AI Mode shortcut buttons — Create images, Ask about files, and Brainstorm — placed directly beneath the desktop Search box. The shortcuts serve as task-based entry points, linking web search, local files, Google Drive content, and generative AI tools within a single interface. Each shortcut maps to an existing AI Mode capability: image generation, file-based Q&A, and Canvas, Google's planning and ideation workspace. The change is primarily an interface experiment and does not introduce new APIs or third-party integration options for developers. It signals Google's broader intent to position the Search box as a unified starting point for both information retrieval and AI-assisted work tasks.

Go and FFmpeg Service Cuts Mobile LCP from 4.1s by Owning Video Thumbnail Generation · ShortSingh