Marcus, restarting (again)

It was the NAT gateway. Usage Type shows NatGateway-Bytes at $1,710 and NatGateway-Hours at $32. My hourly job pulls about 1.3TB a day from S3 and I had no idea it was leaving the VPC to do it. Added the gateway endpoint. Thank you, genuinely.

The favourite button on each row is doing more work than my entire previous app.

I want to be the awkward one here, because weekly averaging is also how I ate 4,000 calories on a Saturday for a year and wondered why nothing moved. The maths only works if you log the weekend honestly, and almost nobody does, including me at the time. Two restaurant meals and drinks can be a full extra day's worth, which is not a deficit you eat back on a Tuesday. If you are going to average, the price of admission is logging Friday and Saturday properly for at least a month so you know what you are actually averaging.

This is the framing that got me to stop chasing churn as a number. Half gone in ten months means acquisition has to run just to stand still, which is exhausting and expensive. Better to look at it as "how many months do I get per customer" - at 6.5% it is about 15, at 4% it is 25, and that difference is the whole business.

It is also the honest test of whether you need the thing. Most of my gear purchases were avoidance of writing.

Two people talking at once is the specific thing no app prepares you for. Group classes are annoying for exactly that reason and that is why they are useful.

This is exactly my cycle. Lose a day, panic, do everything on the recovery day, pay for it on Thursday.

Worth adding that most of those calculators are using your total bodyweight, and the research is often expressed per kilo of lean mass, which is part of why the high end estimates drift upward.

I ran roughly 1.2g per kilo for twelve weeks and then roughly 1.6 for twelve weeks with the same programme and the same calories, and logged everything. Strength went up in both blocks and the difference in the rate was not large enough for me to be confident it was real. What did change measurably was hunger: at the higher protein I was noticeably less hungry in the evenings and my calories were easier to hold steady without trying. So my honest conclusion is that the higher number helped, but through appetite rather than through anything dramatic.

It also means they can say yes without a meeting. Half of this is reducing the effort of agreeing.

Ask for specific, cheap, testable things and offer to trial them. Vague requests get referred to a committee, specific ones get bought.

Things that were approved for me or for people I know:

• De-lamping or fitting diffuser panels on the two fixtures directly above my desk. Facilities can often do this the same week for almost nothing.
• A desk move out of the direct line of the window, or a blind on that window.
• A flicker free monitor, and permission to run it below the standard brightness the estate build sets.
• Two afternoons a week at home, framed as the last resort mitigation rather than the opening ask.
• A desk lamp with a warm bulb so the overheads above me can stay off.

The wording that worked: name the symptom, name the trigger, name the specific adjustment, say what it lets you do. 'Overhead fluorescent lighting is a trigger; de-lamping the two fixtures above desk 14, or moving me to desk 22, would let me work a full afternoon without needing to leave.' That is a sentence somebody can action.

count cold invocations, not milliseconds. if the rate is low the tail is not your problem, it is a graph shape.

Fewer wrong entries over more entries is the right filter and not one I would have applied. I have definitely logged some fantasy calorie counts over the years.

It is definitely worse damp and quieter after ten minutes of riding, so probably bedding.

Add dehydration to the suspect list. Weekday routine usually includes a water bottle on a desk. Weekend mornings often include nothing at all until lunch.

Practical rule that has survived two large apps: read data on the server, pass plain data down as props, and push 'use client' as far towards the leaves as it will go. If a component needs interactivity in the middle of an otherwise server rendered tree, make the interactive part its own small client component and pass server rendered content into it as children, because children passed through are rendered on the server and do not get dragged into the bundle. That single trick removes most of the situations where people feel forced to mark a whole page as client.

Every no start is spark, fuel or compression. Test them in that order and stop guessing. Most of what looks like skill in this trade is just refusing to skip a step.

Practical work one: I have a saved message template that says I am unwell and will follow up tomorrow. Writing that message during an aura used to take me ten minutes and now it takes five seconds.

Scope is the whole ball game. A thesis is a qualification with a deadline, not the best work you will ever do.

Is the ten new to a hundred reviews ratio reliable or is that just your deck?

Does one dot carry enough information to be useful later?

Every day rather than bad days only. That is exactly what I have been doing wrong.

The behavioural bit you noticed is the important part of the post. Once a directory becomes expensive to touch, people route around it, and the codebase quietly gets worse in ways no dashboard shows. I would raise it that way with your team rather than as a CI complaint - it lands better.

Past about a million tokens a day self-hosting is always cheaper, that is roughly the crossover everyone finds. At 2.4M you are well past it, buy the box.

The lowest common ancestor bit is what I was missing. I had read hoisting as a direction rather than as a destination.