mdeeks

We generally try to do things that give a return on investment and don't burn money for no reason. This one most certainly had the ROI. It took 4-ish people to build it in under a year and it removed significant pain and increased developer velocity for hundreds of other developers. Many of whom are working on things that allowed us to be profitable at our IPO.

We're not just a shopping cart product. I thought the same before I first joined but it turns out getting groceries to your door is incredibly hard to do at nation-wide scale.

Right. The devs have full sudo access. They can install and configure whatever they want. Our Bento tool has pre/post setup hooks for all services. So teams can automate this. When something is commonly used we just preinstall it for them. Or some just contribute to the repo. It is all fairly straight forward Ansible and shell scripts. We push nightly to everyone's existing machines.

We don’t use spot because they can be yanked away easily. Savings plans and hibernation are the biggest cost savers.

We actually moved away from containers for dev purposes for almost all of our apps, though many teams still use them. Mainly because they are a bad developer experience. We invested in making apps run consistently in a native way.

Bento was actually our first attempt at doing all of this on our local MacBooks. Many users found that their machine would slow to a crawl and heat up. M1 Macs and now OrbStack instead of Docker Desktop help reduce that. Moving off of docker helped as well. Ec2 gives us the consistent image that is easily testable and repeatable. Egress costs here are negligible.

We dont' keep the machine running all of the time. It stays up running during the users work hours (~8 hours) and then hibernates (the machine is off but memory persisted) which reduces EC2 costs (minus EBS) to zero. We detect activity on their laptop and then wake it when needed.

We at Instacart use them heavily. It is a system my team built called Bento Remote ("Bento" is a local orchestration tool we previously built). We built Bento Remote ourselves because Codespaces was in its infancy at the time and really couldn't meet our needs for things like pre-built images, preserved disks, warm pools, live patching, and being behind our firewall. Bento Remote was and is hugely successful and likely one of the largest boosts to developer productivity in our history.

With Bento Remote get a fresh new fully dedicated EC2 machine in about two minutes from when you run `bento remote create`. It is continuously tested on every merge and verified to function. Everything you need is preinstalled and ready to go. Just connect VSCode to it and start coding. No futzing around getting your local environment to work. Break something on your Bento Remote? Throw it away and get a new machine. Switching projects? Grab a new machine from the pool that is preconfigured for that project. Your settings travel with you.

We had a potentially unique set of circumstances though. Our full stack development requirements were VERY high and needed 64GB to even run our largest and most active apps. This made other 3rd party tools fantastically expensive. Building it ourselves let us fit instacart specific needs and workflows. As well as do clever things like hibernating the EC2 instance after hours when the users laptop is idle. This plus a host of other measures were essential to making this not only cost effective but a large net gain.

So I highly recommend everyone at least take a look and evaluate if you have the need. Start with the SaaS versions and see.

I literally don't understand the first example. What just happened here? Did pkgx see that I failed to run bun and then just like... installed it and ran it for me? Does that mean on my next prompt I should run `bun` or should I run `pkgx` again?

  $ bun
  command not found: bun
  ^^ type `pkgx` to run that
  
  $ pkgx
  running `bun`…
  
  Bun: a fast JavaScript runtime, package manager, bundler and test runner.
  # …

What about the dynamic memory allocation bug?

As the other commenter mentions, you can actually use `docker context` to choose which system you want to use for all of your tools. That docker socket you're referencing is just for the default context if you never chose anything.

Personally, this is the behavior I want and I think that is correct. I just installed OrbStack. That means I want it to take over as the new way to use docker. Switching back to any other system is easy and it is documented here: https://docs.orbstack.dev/faq#revert-docker

I suppose Orbstack should make it more obvious what it did and how to revert though.

Gotcha, totally understand. But I will say that ideas matter far less than the ability to deliver quickly and with high quality. IRC technically has nearly all of the same features as Slack and yet...

So far you're far ahead of the competition in multiple of the ways that matter most. I've tried all of the competition and I don't want any of it on my computer.

The fact that it is dead simple to install, migrate, and use also gives you a huge leg up when we consider adopting it for our company. I absolutely don't want to spin up a support channel to help people migrate to or troubleshoot podman or rancher.

Can you share any long term plans past 1.0? Honestly it feels like a finished product already. Which is amazing for a 1.0!

Probably at this point this is the closest the US can get to declaring war with another major nuclear power. Declaring actual war would probably escalate to a horrific end with nuclear weapons.

The US is explicitly and very loudly supplying a massive amount of weapons to Ukraine for the purpose of killing Russian soldiers and equipment. It feels like a technicality that our troops aren't the ones pulling the triggers.

I'm not weighing in on either side of this argument since I'm undecided, but I think this is what war looks like between nuclear nations.

If this was mine, I'm so sorry :) Just kidding, I actually know my VB6 code is still running in prod.

VB6 did let me learn to program very valuable utilities from scratch, with practically no programming experience. I probably owe my entire career to it and VBA. I took a C++ course once and it nearly turned me off of coding completely. VB6 saved me.

Great question! I don't know. I assume that Tesla pays the owner a flat monthly fee for spots but who knows. Likely some spots are more valuable than others. In some places Superchargers might bring in more people to the local businesses, but in others it might cost the owner limited parking spaces and hurt foot traffic.

Compare that to gasoline at ~$104 to fill a tank of gas on an Audi Q5 (rough equivalent of a Model Y. It has an 18.5G tank and premium gas is $5.65 here). The Q5 does get far more range though. Something like 450+ miles vs the 330 of the Model Y.

Maybe it is better to judge how many miles a dollar gets you? 8 miles per dollar for the Model Y vs 4.3 miles per dollar for the Q5.

If you charge at home during off-peak hours, like most people do, the numbers get a ton better.

It is usually dramatically cheaper to charge at home at night.

For example, in Alameda California I pay:

* 15.17 cents per kWH during off-peak times (nights and weekends)

* 50 cents per kWH during peak times (5pm-9pm)

The Supercharger down the road charges:

* 51 cents per kWH from 12pm-11pm (the time you're most likely to use it)

* 41 cents per kWH from 11pm-4am

* 32 cents per kWH from 4am-12pm

No one is going to leave their car at a Supercharger overnight though. So more realistically you're paying over 3x the cost if you use public Superchargers.

My Model Y has a 75kWH battery so its about ~$11 to charge it from zero to full at home vs about ~$38 at a Supercharger during the day.

Anecdata: It directly lead us (Instacart) to try and then adopt Tailscale. Many of us had used it at home and were happy nerds. This gave it a huge initial leg up vs other "enterprisey" VPNs when we were in the evaluation stage.

Tailscale sold itself after that. The docs were excellent and it really is simple to use and run. I was able to do a full PoC in day and prove that I could join all of our environments and clouds into one VPN and have DNS resolving correctly everywhere.

Sorry this article was EV specific but I see now the OP meant electricity costs in general

For PG&E it appears they have an EV specific rate plan that is cheaper. It says it is $0.27/kWh for off peak charging.

https://www.pge.com/pge_global/local/images/data/en-us/rate-...

I can't link directly to the tab but its the second tab on this page: https://www.pge.com/en_US/residential/rate-plans/rate-plan-o...

If you're lucky enough to live in Alameda, it has its own municipal power. Their EV rates are significantly cheaper at $0.15/kWh https://www.alamedamp.com/393/TOU-Time-of-Use-Rate-for-EV-Ow...

Do you have a source? That is really interesting if so! I'll plug my car in more often then.

I remember seeing Elon Musk say that our Model 3 was rated for 1500 recharge cycles. Our warranty is for 70% capacity after 8 years or 120k miles. That hints to me that cycling is the main cause of degradation.

https://www.tesla.com/support/vehicle-warranty

Caddy is great! But has the absolutely insane default of logging all headers. Make sure you change that before you log auth headers

Chrome used to be a lot faster years back so I switched. Firefox sadly has never given me a reason to switch back. Chrome is still fast and generally easy to use. (Maybe others here can give me good reasons to switch?)

Chrome also has some obscure features I rely on that Firefox does not have (or just doesn't do well):

  * Firefox does not support Touch ID on macOS. I can't actually log into a lot of my sites

  * I can't create a desktop app version of a site (Chrome > Tools > Create Shortcut as window). I use this to put google meet, jira, and some other internal tools as dedicated dock apps.

  * Chrome lets you run a work and personal profile in different windows. It's just a cmd+` away. It also remembers which profile you had viewed last and opens links in the right one. I didn't like multi-account containers in Firefox because I always had to select the right profile when I click a link.

There are a bunch of videos of it on twitter right now. Seems confirmed. https://twitter.com/saintjavelin/status/1680826559626706944?...

OrbStack is shockingly good. We have tools that are deeply integrated with Docker Desktop and the CLI and it all worked perfectly out of the box when I switched to OrbStack. It also takes like seconds to start and has yet to overheat my laptop or drain my battery.

Highly recommended: https://orbstack.dev/

Mainly because Tailscale is popular with the HN crowd. Many people use it for their personal home networks, and some companies do to (we do at Instacart).

The nice thing about this is it means you can invite anyone into your network without requiring them to have a Google/Apple/Github account. They just need a phone. It also means you can be much more confident inviting these people because they are unlikely to get phished due to a crappy password.

One of the common complaints in HN comments about Tailscale is they don't want to trust Google/Apple/Github/etc for various reasons. This removes that need.

Hm I have quite the opposite experience. I've never had so many bugs with Twitter as I do right now.

The feed random refresh bug never went away and somehow managed to get worse. Right as I'm reading something it'll vanish without a trace. I probably see this five times per day.

Sometimes I'll find my app on the "Following" feed even though I never actually switched to it (I'm weird, I actually enjoy the For You feed a lot).

Sometimes the content clearly hasn't refreshed because I see new things on my desktop that aren't on my mobile app.

The quality of the content has taken a sharp nose dive too. Almost everything is useless threads from "thought influencers", memes, or completely random posts that I have never interacted with ever in my life (anime memes? a thousand posts about the barbie movie)

The quality of the replies is awful now that blue check users are all at the top. It essentially made it "pay to be seen" rather than having the more interesting replies at the top. This makes me the most sad I think :(

Aside from that though, it feels like roughly the same app but with some small tweaks here and there.

With their default model it did extremely badly with my question to build this script. It wasn't even close to right.

When I changed it to "Use Best Model" it gave a very good response but I'm pretty certain it was just piped to OpenAI GPT-4.

My prompt:

Write a python script for me that deletes AMIs from AWS that I own which have the tags role=bento-remote and bento_image_type=profile. It should bucket the AMIs by another tag called profile. It should delete images older than 14 days but always keep the 10 most recent even if they are older than 14 days. It should print the name of the AMI it is about to delete. It should run the deletes in parallel so it is fast.

This is significantly worse than ChatGPT with GPT-4 for coding. I had an example from last week where I wanted to write a script to delete excess AWS AMIs based on some tags of ours. Then modified it to bucket them by a tag value and keep the last ten of each bucket. Then parallelize it so it is fast (we had 50k of them, oops).

ChatGPT was able to completely write this perfectly. Absolutely mind blowing. Saved me probably 1-2 hours of looking things up and fiddling with it.

Bard failed multiple times in multiple places. It only found AMIs that exactly, down to the second, were two weeks old. The filter json looks wrong ('Key' isn't valid AFAIK). It didn't bucket them and just kept the newest 10.

In either case both of these are mind blowing. If ChatGPT didn't exist, Bard would still be very helpful at getting me started. ChatGPT did it all without me making any changes though.

ChatGPT log: https://gist.github.com/mdeeks/de297f1bc8cbd00fe2db01e6232aa...

Bard log: https://gist.github.com/mdeeks/1404e09da8879b94469166927ddff...

I love being in the city and working side by side with my coworkers. But I absolutely do not want to commute two hours daily. I have kids, so I must be home by 5pm and also must help get them ready for drop off at 8am. That leaves no time for work + commuting.

Mass transit in the bay area sucks. Bart is packed, loud, and uncomfortable. I feel unsafe at my station (Fruitvale). So my only other option is to take the bus which is an hour each way because of traffic.

We (Instacart) were looking for alternatives before it started costing money. One of the primary reasons we pivoted and built remote development environments was because of the horrible performance of Docker Desktop. Even when running containers that do nothing I'd find my CPU was high for no reason. Anyone on an Intel Mac would essentially have a loud lap warmer with 2 hour battery life.

We are actively looking for alternatives that a) has minimal performance impact on our devs laptops b) is stable (We have most of our devs locked onto a specific version of Docker Desktop because latest was randomly crashing).

If the pricing is competitive we will likely be customers when you're out of beta.

Atlantis is basically just a terraform oriented command runner. I believe you can create a custom Atlantis workflow that'll just run `pulumi preview` or `pulumi up`. That's how Atlantis supports terragrunt as well: https://www.runatlantis.io/docs/custom-workflows.html#terrag...

Our experience with Terraform is exactly the same as what you're describing with Pulumi. We have hundreds of stacks and they are all implemented differently because they are written by a hundred different people with varying knowledge of Terraform. Terraform is a goofy language and it's very hard to do even simple things like conditionals. There is also a massive amount of copy/paste because people have no idea how to set up their provider or remote states and terraform has no real way to make those DRY.