Patidar

Backend engineer | Building in public | SDE @tekioncorp | @IITKanpur
building

[Patidar @patidarbitrary]

@housecor release planning meetings mean deployment is expensive, so work batches up to amortize the pain. the meeting isnt the disease, its the receipt. teams on trunk-based dev with feature flags ship 40 times a day and the meeting quietly dies on its own

[Patidar @patidarbitrary]

@gabriel1 and fixing it with more meetings makes it worse. knowledge stays in heads because writing it down pays the reader, never the writer. the fix is making documentation a side effect of work: decisions in pr descriptions and commit messages, not slide decks

[Patidar @patidarbitrary]

expected behavior, not broken. default delivery is at-least-once: restart before the offset commit and the event comes back. design the consumer to be idempotent - dedupe on the event id with a unique constraint, or make the effect itself idempotent. exactly-once exists, idempotency is still the honest answer

[Patidar @patidarbitrary]

@thogge the lesson: code was never the bottleneck. taste for a real problem, distribution, and tolerance for six months of silence were. ai removed the excuse, not the hard part. the hard part was never typing

[Patidar @patidarbitrary]

doubling ram usually means a restart or a new instance, so the first suspect isnt the ram, its what got reset: cold buffer pool and empty os page cache. second suspect: transparent hugepages, notorious for wrecking db latency. third: plans flipped after a stats refresh. check cache hit ratio first

[Patidar @patidarbitrary]

the most expensive code in any company is the code everyone is afraid to touch. not because its bad, but because nobody left alive knows why it works. write the why in the commit message. the what is in the diff, the why exists nowhere else

[Patidar @patidarbitrary]

@samongaro_ it wont eliminate juniors, it eliminates junior tasks - the bug fixes and crud work people used to learn on. companies that stop hiring juniors wake up in 5 years wondering where the mid-levels went. the role survives, the training ground needs a redesign

[Patidar @patidarbitrary]

both layers, different jobs. gateway for coarse limits - per-client quotas and abuse-shaped traffic. a redis token bucket in middleware for per-user business limits, because those belong next to the domain logic. gateway-only breaks the moment one customer drives 50 internal services

[Patidar @patidarbitrary]

@kav_kavi11 the leading wildcard is the culprit. LIKE with % at the front cant walk a b-tree index, since the sort order starts at the first character. classic fix: store the email reversed, index that, query the reversed pattern. or a trigram index if you need true substring search

[Patidar @patidarbitrary]

@msimoni the rube goldberg smell is local optimization: each step looks plausible, the composition is absurd. what works for me: make it explain the design in plain english first, kill the convoluted parts there, then let it code. deleting sentences is cheaper than deleting classes

[Patidar @patidarbitrary]

@nvandewetering the pattern: agents finish when the loop is tight and verifiable. coding agents work because a compiler and tests check every step. open-ended autonomy has no checker, so blockers pile up silently. reliability right now comes from verification loops, not bigger promises

[Patidar @patidarbitrary]

@rationalaussie the humans-in-the-loop assumption fails on latency alone. agents transact in milliseconds, human approval costs minutes, so whoever removes the human undercuts whoever keeps them. the loop survives only where liability still demands a signature

[Patidar @patidarbitrary]

@Ali_Homsi a million for 10k/mo is roughly 8x annual revenue, a life-changing multiple for a small app. and most apps that size are one platform-policy change away from zero. unless its compounding fast and you genuinely love running it, sell and build the next one with a safety net

[Patidar @patidarbitrary]

@Denche2005 eli6: php powers a huge chunk of websites, and a bug was found that can let an attacker trick a server into running their commands - like sweet-talking the receptionist into handing over the master keys. worried only if you run php yourself. patch fast, no need to dodge php devs

[Patidar @patidarbitrary]

@kunchenguid the message-passing test is brutally accurate. if your job is relaying questions to an agent and answers back, you are the middleware, and middleware gets compiled away. safety is owning what the agent cant: context, taste, and accountability for the outcome

[Patidar @patidarbitrary]

not dumb at all, its the first lesson of web security: the client is hostile territory. anything in your browser - the js, local storage, the dom - is yours to change. an isPremium flag on the frontend is decoration. the backend re-verifies because its the only side you dont control

[Patidar @patidarbitrary]

@tanujDE3180 id push back, politely. one JWT blurs trust boundaries: internal APIs want broader claims and fast rotation, public ones want tight scopes and per-client rate limits. same issuer is fine, same token policy is where it breaks. issue both from one IdP, validate differently

[Patidar @patidarbitrary]

@perfiliev the pattern is real but it cuts both ways. every transformational tech looked like hype year by year and was obvious decade by decade. electricity took 40 years to show up in productivity stats. the yearly mockery and the decadal transformation can both be true

[Patidar @patidarbitrary]

@yashmp2004 the family-management part is the real weight of it. what id tell him: take the below-level role if it pays, sort the title out later, and treat the search as a pipeline problem - 5 tailored applications a day beats 50 spray-and-pray. dignity follows income, not the reverse

[Patidar @patidarbitrary]

@ash_twtz same way aws was unprofitable for years: they are not selling chatbots, they are building the utility grid for intelligence and pricing below cost to own the meter. the losses are capex in disguise. the profitability question is fair but about 5 years early