Method and coverage

10,332 posts read into problems so far, 913 of them corroborated. 184 sources read; 10% of registry companies have at least one problem.

How a post becomes a problem

1ReadEvery day at 05:25 UTC (again at 17:25 UTC only if that run failed) the scan reads public sources: Hacker News through the Algolia API, public Reddit feeds (a few subreddits a run, slowly), Stack Overflow, the most-reacted GitHub issues of major data and AI repos, Discourse forums, Bluesky and Mastodon public posts, YouTube video titles, 1 and 2 star App Store reviews, and outages, deprecations, breaches and advisories from other Glyph Ibis sources.
2Keep what states a painA post counts when it says something hurts: "struggling with", "doesn't work", "workaround", "is there a way", "rate limit", "nightmare" and about a hundred more phrases. Issues, unanswered questions, forum help topics and low-star reviews count by nature. On general sources the post must also be about tech.
3Name the issue and the companiesEach post is matched to a known issue (the table below) and to the companies and products it names: a product list plus every company in the companies.fru.dev registry, matched by exact name.
4GroupPosts with the same issue about the same product become one problem once two posts say it; until then they join the general problem. Posts about a product with no known issue are grouped by the title word they share (3 or more posts). Statements are written by these rules, 12 words or fewer.
5MeasurePer problem: posts per day, the last 7 days against the 7 before (rising, steady, fading, new), days active, distinct people and places, frustration from strong words, and the top three links. Persistent means 5 or more active days across two weeks and still active this week.
6Keep historyPosts, day counts and daily snapshots are append-only; every change is on the Changes page. A cluster backed by fewer than 3 posts from 2 people or places stays marked unverified.

Coverage by source

Source	Problems	Posts
GitHub issues	1,318	6,204
Reddit	468	1,331
Glyph Ibis	81	908
App Store reviews	113	697
Community forums	231	684
Hacker News	122	289
Stack Overflow	108	173
Mastodon	23	29
Bluesky	7	9
DEV	7	7
Lobsters	1	1

Registry: 2,769 companies.fru.dev companies enrolled, 268 with at least one problem (10%). The rest are watched on every run. Last run: daily, ok, 232 posts added, 45 new problems, 0 new sources.

Gaps, named

Reddit answers scripts slowly and blocks bursts, so each run reads a few subreddits in turn; a subreddit is read about every five days.
Slack: only communities that publish public archives. Private workspaces are never joined or read.
YouTube comments need a YouTube Data API key; until one is set only video titles are read.
X, LinkedIn, G2 and TrustRadius are not read: they need a login or do not allow it.
Bluesky is read by search, Mastodon by public tags on mastodon.social only.
Rules miss sarcasm and can file a post under the wrong issue; clusters with few posts are marked unverified. Suggest a fix on any problem page.

The issues the rules know

Issue	Category
LLMs give confidently wrong answers	LLM apps
Prompt injection makes AI features unsafe to ship	Security
API rate limits and usage caps block real work	LLM apps
Context windows run out on real workloads	LLM apps
Model quality silently gets worse between versions	LLM apps
AI agents are unreliable, loop or stall midway	LLM apps
Models fail to return valid structured output	LLM apps
RAG retrieval returns irrelevant or missing context	LLM apps
No reliable way to evaluate LLM app quality	LLM apps
Prompts break or get ignored when models change	LLM apps
LLM responses are too slow for users	LLM apps
LLM token costs are hard to predict and control	Cloud and tool cost
Unclear what AI vendors do with company data	Security
Running LLMs locally hits memory and speed limits	MLOps
Fine-tuning is costly and results are unpredictable	MLOps
GPU access, drivers and CUDA versions are painful	MLOps
Deploying and serving models is harder than training	MLOps
ML experiments are hard to reproduce and track	MLOps
Model and feature drift goes unnoticed in production	MLOps
AI-generated code is buggy and hard to review	AI coding
AI assistants lose context in large codebases	AI coding
AI tool plans, limits and pricing keep changing	AI coding
Upstream schema changes break downstream pipelines	Data engineering
Bad data slips into production unnoticed	Data engineering
Backfills are slow, costly and risky	Data engineering
Data pipelines fail silently or break often	Data engineering
Debugging orchestration and DAG failures is hard	Data engineering
Ingestion connectors and CDC are fragile	Data engineering
Streaming pipelines lag, drop or reorder events	Data engineering
Data modeling choices are unclear and hard to change	Data engineering
Hard to trace lineage and find trusted tables	Data engineering
Model, table and pipeline sprawl becomes unmanageable	Data engineering
Testing data pipelines and SQL is hard	Data engineering
Small files, skew and partitioning hurt performance	Data engineering
Databases hit connection, locking and scaling limits	Databases
Vector search quality and scaling are hard to get right	Databases
Queries run too slowly and are hard to tune	Databases
Teams do not trust or agree on the numbers	Analytics and BI
Dashboards are slow, broken or unused	Analytics and BI
Endless ad hoc requests swamp data teams	Analytics and BI
Critical work still runs on fragile spreadsheets	Analytics and BI
Warehouse compute costs spiral out of control	Cloud and tool cost
Observability and logging costs outgrow the value	Cloud and tool cost
Cloud bills are surprising and hard to attribute	Cloud and tool cost
Tool prices rise faster than the value they give	Cloud and tool cost
Kubernetes is too complex for most teams	DevOps and infra
CI builds are slow and tests are flaky	DevOps and infra
Infrastructure code drifts from what is really deployed	DevOps and infra
Alert noise and on-call load burn teams out	DevOps and infra
Permissions and IAM are confusing and error-prone	DevOps and infra
Secrets and API keys leak or are hard to rotate	Security
Containers are slow to build, bloated or crash-looping	DevOps and infra
Jobs and services run out of memory	DevOps and infra
Malicious packages slip into dependencies	Security
Vulnerability scanners flood teams with noise	Security
Breaches expose customer and company data	Security
Login, SSO and OAuth flows break	Dev tooling
Service outages and slowdowns disrupt work	Outages and change
Deprecations force unplanned migrations	Outages and change
Upgrades and new releases break working setups	Outages and change
Migrating between platforms is slow and risky	Data engineering
Vendor lock-in makes switching tools costly	Cloud and tool cost
Documentation is outdated, thin or confusing	Dev tooling
Vendor support is slow or unhelpful	Dev tooling
Dependency conflicts break installs and builds	Dev tooling
Integrations between tools are brittle glue code	Dev tooling
Local development setups are hard to get right	Dev tooling
Legacy systems and tech debt slow everything down	Dev tooling
Tools are clunky and unintuitive to use	Dev tooling
Everyday tools feel slow and laggy	Dev tooling
Error messages are cryptic and unhelpful	Dev tooling
Worry that AI is replacing engineering jobs	Careers
Leadership pushes AI mandates without clear value	Careers
Tech job hunting is brutal right now	Careers
Technical interviews feel broken and exhausting	Careers
Burnout and overwork across tech teams	Careers
Hard to keep up with the pace of new tools	Careers