Story of how agent failed

We have our own automation tool, sort of airflow. I needed to change some stuff in the cloud. We have crazy complicated way of doing that, looks more like what I have on the picture.

I was prompting (grok 4.7, cursor) , asking to check our docs, slack, etc and give me a way how to do a change, test/validate it, and be sure it will work.

It gave me good looking but not working plan all the time, no proper test plan, missed release cycles etc.

I talked to my colleague who had hands on exp with that - in 5 mins I got the plan how to do it which is SOOO different from what AI gave me.

#ai #aifailed

Maybe job security move is to work on irreplaceable very bad internal system which no one knows how it works?

Faangs typically have a lot of such things

Will appreciate your stories of how ai failed for you in your use case

Daily 432nd day streak

Pretty easy, but not some real pattern here.

https://leetcode.com/problems/subarrays-distinct-element-sum-of-squares-i/description/?envType=daily-question&envId=2026-10-15

#daily #easy #nopattern

I have seen this q in Threads:

Tech interview questions | Database

The orders table has ~50 million rows. You need to add a new column is_gift with a default value of false.

A developer ran the migration, and after 10 seconds the entire service stopped responding, even for reading orders. The migration process had to be killed.
How should this migration be done correctly?

The expected answer probably implies some db knowledge, potentially avoiding index creation, and potentially going with NULL first.

BUT

The actual way of doing it if you respect customers, your engineers and oncall - is typical 3 step migration:

Bring new storage with new field. Hydrate it with existing data

1. Write to both, read from first storage
2. Write to both, read from second storage
3. Write and read from second storage (that has flag)


Worked on 1M rps. Applies regardless of storage type.

It must be done with feature flags IN THE CODE.

This way you control where and when you enable migration step, you have monitoring, and THE MOST IMPORTANT: you can rollback any time

Good luck with that.

Actually number of records is irrelevant. It is about reliability.

You can never be sure your migration will work without error 100% on prod instance.

If it does not - you have downtime, lost data(as recovery is from backup, or roll forward the fix, that might not work as well).

You dont have control and monitoring. In sharded env - again, you cannot do it atomically on multiple shards.

Aaand it does not work for all types of migration/storages (it can be much more complicated than boolean field)

Especially on highload systems - never do that. Always do 3 steps approach

This is golden rule we have migrated tons of different shit at MS like that. No principal would approve any other approach.

Yes, let me correct myself.
Volodymyr is right about that specific case.

Probably nothing will happen and it will work on single db(with low rps)

I would go that way ONLY
- if I can afford downtime + backup + losing data
- If I am okay to be down until I hot fox and roll forward

Otherwise - 3 steps.

I have trust issues with software, sorry

After I worked some years

I just hope software for healthcare and airplanes along with defence area is built differently

Would you come to me as to doctor, if I am same quality doc as I am engineer?😆😆😆😆😆

Daily 433d day streak.

Same problem as yesterday, but bigger complexity.
It is segment tree problem, suggest to skip, if you want to train segment tree - try simpler one (in reply to that message).

https://leetcode.com/problems/subarrays-distinct-element-sum-of-squares-ii/description/?envType=daily-question&envId=2026-10-16

#daily #hard #pattern

Someone has sent me.

The problem is that similar thoughts I have heard from ex-Google director, currently Principal at Snow….

The guy is from Moldova. We had a Leetcode mock interview in the beginning of the summer.

Now he relocated to Ireland, interviewing with many companies, and eventually we met in person in noname coworking, cause Snowflake does not have nice office to show😆

Let’s wish him offers, as market in Ireland at least is not that great.

So was it 2026 instead of 2015?)

THIS IS SO HILARIOUS😆😆😆😆

Daily 434th day streak

Interesting and unique problem.
Was pretty hard for me to build an intuition. But solution has pattern, so I would recommend.

https://leetcode.com/problems/minimum-swaps-to-sort-by-digit-sum/description/?envType=daily-question&envId=2026-10-17

#daily #medium #pattern

Had to show my CT scan to Ukrainian friend. Wanted to make it as easy as possible to see that scan so no need to install any app, etc

This is why I love about AI. 15 mins + my manual deployment: here you go, you have CT scan viewer. Just like that. Everyone can do it. Cheap and easy. Scary... and good at the same time

Btw look how big my heart is😁 I thought it is smaller.