Tags

Rather than modelling your data model based on what it is (relations, etc.). Model based on how the application will need to query it (key value, top 50, complex search, etc.) Optimise…

Server operation time + network overhead. For example, Redis operations are quite fast (couple of ms), there will always be network overhead.

One way to do AI Multi Agent could be to route the request to the right model depending on the complexity needed, at least to save costs. https://blog.bytebytego.com/p/how-smart-model-routing-can-cut-llm.

Reviewing code is a skill that won't go away. No shortcut to it, get good the old school way. Take the time to write the project description and code structure yourself, it…

Frontier models are good by themselves (ie Claude Code) Single agent (could) need more human in the loop to review, where multi agent could review itself.

Smolagents vs langgraph. To test the output of your models, you can rely on both deterministic and non-deterministic guardrails (LLM as Judge). Remember that AI outcome is not deterministic so it's safer…

Maybe not straight user ID in api. Jwt json web token contains signed json data about the user, like what role they have. Every request contains a signature that the server can…

Failing fast is better than failing slow. Circuit breakers allows not calling services that are broken so that they're not overloaded. Half open allows letting a few requests through to see if…

Api gateways can combine results from different apis into one. Routing: https://www.hellointerview.com/learn/system-design/deep-dives/api-gateway.

Django rate limiter? In memory per server. You don't get a global picture tho. Or separate service but extra latency. Most common is to have it with the api gateway. Rate limiting…

Good to handle the odd error. Beware of overloading already failing services with even more load. API Circuit breakers will help.

Set long timeouts intentionally, services should be fast, unless it is purposefully (can't do otherwise) slow.

Availability patterns. There are two main patterns to support high availability: fail-over and replication. Fail-over. Active-passive. With active-passive fail-over, heartbeats are sent between the active and the passive server on standby. If…

Separates compute from storage. Leverages s3 as elastic storage, making it scale way beyond what traditional dbs can do. For example, for PostgreSQL it starts struggling around 100GB to 1TB. With aurora,…

In old existing product data, it was stream processing, processing one spider item at a time. whereas in dogs, it was batch processing, processing 'big' sqlite files.

Cache = trade off speed vs freshness. You can also use the CDN as a cache for api results. Short term. Ok if only a few. Consider how often it changes, how…

Cache aside: app checks cache and fetches from DB on miss then updates cache. Read through: app checks cache, the cache will automatically fetch from the DB if missing. The difference between…

Cassandra is masterless. Cassandra has no Database ACID support. Bias towards Availability. Write heavy. Zero downtime needed (in traditional master/slave, there is a small downtime if master is down for the failover…

Spread out servers that store copies of content. Page load will be faster as there will be a server closer to the user's location, usually. These usually store and cache content like…

make adr folder. make agents.md and tell claude.md to read that to be ai agnostic. write preferences to agents.md, ie testing pure functions or disallow running certain commands. but keep it light…

What did I build? Why did I structure it this way? Why this database? What assumptions did I make? What are the failure cases? What did I test? What doesn't it handle…

Be comfortable with: dictionaries / sets / lists. comprehensions. iterators/generators. collections. sorting. exceptions. context managers. dataclasses. typing. JSON. HTTP. files. async basics. or. descending: Remember the distinction: creates the entire list. Whereas:

Database Replication and read replicas. Something something geolocalisation. Postgres has extensions for this.

Connection pooling allows avoiding TCP handshake overhead. Keep some connections ahead.

Weak consistency. After a write, reads may or may not see it. A best effort approach is taken. This approach is seen in systems such as memcached. Weak consistency works well in…

Consistency = every read gets most recent write. Availability = the system responds even if it's not the most recent data. For example with Ticketmaster, consistency is way more important, but for…

Depending on the data you need to fix, if it's simply releasing something, then a TTL on a row is better.

Removes the need for a slow sequential scan. B-tree is the most popular, allows range queries with almost O(1). Hash index (key mapping) is definitely fast and O(1) but can't do range…

Synchronous is slower, but ensures stronger consistency.

Write-heavy: Too much data for 1 instance (storage or write per second) Read-heavy: when read replicas or caching are not enough. Global queries will be slower. Complexity to operate and keep data…

Distributed configuration management, zookeeper is a bit outdated but good for this. Example: The API Gateway keeps an open connection to config server and pushes changes straight away. Other not as good…

2 phase commit and saga patterns are the most common ways to solve this. You can have a coordinator plan across services and commit all at once, but this is fragile if…

Amazon DynamoDB and Amazon DocumentDB are both fully managed NoSQL databases on AWS, but they are built for entirely different data models and architectural needs. Choose DynamoDB if you need predictable, single-digit…

It builds an inverted index to make searching documents by terms quickly. Example: taylor: (event1, event2) Tokenization in Elasticsearch is the process of breaking a stream of text characters down into smaller,…

Dealing with important events that needs to be stored forever (ie refund events) When you need replayability of events (reprocess historical data) Real-time data processing.

Forward proxy: next to the client before reaching internet. Reverse proxy: server side to go to the right server. API gateway: reverse proxy that does more than routing. it handles auth, rate…

What worked for others in other scenarios might not work for you. Concrete system design discussions are best done by the people involved in working in the system. It'll be more specific.…

This is good for when clients need to choose which fields to use and make complex joins in 1 query. With graphql, it's easy to introduce the n+1 query problem where you…

Horizontal scaling only works if servers are stateless aka, server needs to treat the request as if the user came in here for the very first time.

Good foundation is still key. AI is a signal amplifier, not just a tool. It will amplify the good and the bad. System design is still very important (some say that hyper-generalists…

We still think in black boxes and inputs/outputs like before. Collaboration is still key (in theory). We still think in processes rather than language syntax. Systems thinking. Refactoring. Engineering with ai is…

Title is misleading, there is no "should". The best managers are not removed from the craft for too long.

1KB ish. Some common fields: But what takes most space is database overhead, index, page, etc.

Treat AI as a slow and non-deterministic (AI outcome is not deterministic) worker that can get things wrong, don't use it for critical paths like auth. A good chef that delegates is…

PostgreSQL. Strengths: transactions, joins, constraints, flexible queries. Weaknesses: horizontal scale/sharding complexity, not ideal for huge analytical scans. DynamoDB. Strengths: massive scale, predictable low-latency key access, operational simplicity. Weaknesses: access-pattern-driven modelling, limited relational…

Code is cheap now. Good manually crafted code is harder to stand out in the flood of AI slop. But now everyone can generate lots of code.

Technical debt is visible but comprehension debt is not. Don't ship it if you don't understand it. Code is fast, but slow down to keep a good mental model if necessary.

Client generates an ID to send to the server that is unique when making operations. In case of failures or retries, this ensures operations on the server side are done only once.

An image or a song would be immutable data. The access pattern for those is a lot of reads, very few writes. These would be best stored in a system where it's…

If junior developers don't experiment and learn from getting things wrong, they won't develop judgement and become senior. It's important that AI doesn't replace your thinking. Juniors feel more empowered, 50% velocity…

In Youtube - Yannick Dalbin – Ce que 1 an à laisser l'IA coder m'a appris sur le métier de développeur, he says that if there's an issue with what the model…

Can they do the job? Not necessarily right, now but you need to know that they can learn how to do it. Unless you have constraints that they need to be performant…

Kafka wins when: You need persistence. Strict ordering at scale. Stream processing. SQS+SNS wins when: You're ok with the AWS structure and minimal operations. Retry + DLQ built-in. Ordering does not matter…

useful for production apps that leverage llms beyond calling openai/claude with an api and show the results. this is useful if you need things like subagents doing different things (searching, etc.), orchestration,…

Latency is the time to perform some action or to produce some result.

https://www.techinterviewhandbook.org/grind75/.

Use AI for code reviews when the risk is low.

Single point of contact for user requests. It does health checks for all servers to know if some have issues. Distribution: Round Robin: simple and sequential. IP hashing: allows the same user…

Session database. django-session. Having a session database rather than storing in the server's memory, allows avoiding sticky sessions. It's best to have sessions expire so that the data gets cleaned eventually. When…

operations too long for a live server, get feedback to users quickly. reduce load on a live server. when you want the ability to scale differently for different tasks (ie retrievers need…

read: 20k queries per second. write: 10k writes per second.

When I'd use DDB: if key value access is enough. I need heavy write (more than 10k writes per second) automatic and managed sharding. What you're giving up with postgres: relations and…

MySQL is thread per connection, where PostgreSQL is process per connection. MySQL is seemingly faster for lots of concurrent simple reads. PostgreSQL takes up more memory, but handles complex queries better with…

Pre signed urls rather than streaming through api servers with bigger timeouts. This frees servers to do other things, this is supposedly a secure way to get clients direct access to blob…

A model has to: find every one of those. realise they're related. ignore 99% of everything else. remember them all. compare them. That's basically Retrieval-Augmented Generation (RAG), except your retrieval system is…

Redis is great for data that needs to be accessed by key quickly. Especially if that data is temporary with the TTL feature, like a cache a ticket locking system. It would…

Memcached is supposedly a bit faster as it is multi threaded. Redis has better support of more data types. Redis has optional disk persistence.

REST favours plural nouns. Use path parameters for mandatory parameters, and use query parameters for optional ones. Rest is good for public facing Apis, easy to reason about. Http for external is…

When is GraphQL is a good idea? When flexibility of what data to retrieve is important, but it is possibly rarely worth the overhead.

A reverse proxy is in front of a server. Intermediary so that clients never talk directly to servers.

An API gateway is a specialized type of reverse proxy. The easiest way to differentiate them is by their core focus: a load balancer focuses on scale, a reverse proxy focuses on…

Is it basically doing Role - Software Engineer in the real business world?

RPC tends to be more for internal service communication. REST has a bit more overhead with http. Rpc tends to use more efficient formats like protobuf and grpc rather than http. This…

Either each service triggers the next one, or there is an orchestrator managing the whole thing. Queen was an orchestrator whereas in PIMS, it was choreography style, retriever triggers transformer, etc.

A service is scalable if it results in increased performance in a manner proportional to resources added. Generally, increasing performance means serving more units of work, but it can also be to…

It usually builds an inverted index for fast querying, like: Tokenization: Tokenization is the process of breaking a piece of text into individual words. This allows you to map from words to…

Postgres is fine to begin with for basic search. Select where like term. When would somethint else be more relevant ? Wildcard search are slow tho, it'll scan the whole db. But…

Sticky session means it'll only work if the user has time come back to a specific server. Stateless means server needs to treat the request as if the user came in here…

For example between search and main db. When introducing another database, they need to be synced somehow. Easiest: write to both in app code (what if it fails and also makes it…

Hello interview in a hurry is the best resource: https://www.hellointerview.com/learn/system-design/in-a-hurry/introduction. Book - Designing Data-Intensive Applications (haven't extracted all notes yet)

https://github.com/donnemartin/system-design-primer. They have Anki flashcards too.

Throughput is the number of such actions or results per unit of time. Generally, you should aim for maximal throughput with acceptable latency.

WebSockets is 2 way open connection. SSE (Server Sent Events) is one way. API usually have a timeout of 30-60 seconds, configured by an outside system, often. If the operation you're doing…

Server operation time + network overhead. For example, Redis operations are quite fast (couple of ms), there will always be network overhead.

Failing fast is better than failing slow. Circuit breakers allows not calling services that are broken so that they're not overloaded. Half open allows letting a few requests through to see if…

Api gateways can combine results from different apis into one. Routing: https://www.hellointerview.com/learn/system-design/deep-dives/api-gateway.

Django rate limiter? In memory per server. You don't get a global picture tho. Or separate service but extra latency. Most common is to have it with the api gateway. Rate limiting…

Good to handle the odd error. Beware of overloading already failing services with even more load. API Circuit breakers will help.

Set long timeouts intentionally, services should be fast, unless it is purposefully (can't do otherwise) slow.

Availability patterns. There are two main patterns to support high availability: fail-over and replication. Fail-over. Active-passive. With active-passive fail-over, heartbeats are sent between the active and the passive server on standby. If…

In old existing product data, it was stream processing, processing one spider item at a time. whereas in dogs, it was batch processing, processing 'big' sqlite files.

Cache = trade off speed vs freshness. You can also use the CDN as a cache for api results. Short term. Ok if only a few. Consider how often it changes, how…

Spread out servers that store copies of content. Page load will be faster as there will be a server closer to the user's location, usually. These usually store and cache content like…

Database Replication and read replicas. Something something geolocalisation. Postgres has extensions for this.

Connection pooling allows avoiding TCP handshake overhead. Keep some connections ahead.

Weak consistency. After a write, reads may or may not see it. A best effort approach is taken. This approach is seen in systems such as memcached. Weak consistency works well in…

Consistency = every read gets most recent write. Availability = the system responds even if it's not the most recent data. For example with Ticketmaster, consistency is way more important, but for…

Depending on the data you need to fix, if it's simply releasing something, then a TTL on a row is better.

Distributed configuration management, zookeeper is a bit outdated but good for this. Example: The API Gateway keeps an open connection to config server and pushes changes straight away. Other not as good…

Dealing with important events that needs to be stored forever (ie refund events) When you need replayability of events (reprocess historical data) Real-time data processing.

Forward proxy: next to the client before reaching internet. Reverse proxy: server side to go to the right server. API gateway: reverse proxy that does more than routing. it handles auth, rate…

What worked for others in other scenarios might not work for you. Concrete system design discussions are best done by the people involved in working in the system. It'll be more specific.…

This is good for when clients need to choose which fields to use and make complex joins in 1 query. With graphql, it's easy to introduce the n+1 query problem where you…

Horizontal scaling only works if servers are stateless aka, server needs to treat the request as if the user came in here for the very first time.

Client generates an ID to send to the server that is unique when making operations. In case of failures or retries, this ensures operations on the server side are done only once.

Kafka wins when: You need persistence. Strict ordering at scale. Stream processing. SQS+SNS wins when: You're ok with the AWS structure and minimal operations. Retry + DLQ built-in. Ordering does not matter…

Latency is the time to perform some action or to produce some result.

Single point of contact for user requests. It does health checks for all servers to know if some have issues. Distribution: Round Robin: simple and sequential. IP hashing: allows the same user…

Session database. django-session. Having a session database rather than storing in the server's memory, allows avoiding sticky sessions. It's best to have sessions expire so that the data gets cleaned eventually. When…

operations too long for a live server, get feedback to users quickly. reduce load on a live server. when you want the ability to scale differently for different tasks (ie retrievers need…

Pre signed urls rather than streaming through api servers with bigger timeouts. This frees servers to do other things, this is supposedly a secure way to get clients direct access to blob…

REST favours plural nouns. Use path parameters for mandatory parameters, and use query parameters for optional ones. Rest is good for public facing Apis, easy to reason about. Http for external is…

When is GraphQL is a good idea? When flexibility of what data to retrieve is important, but it is possibly rarely worth the overhead.

A reverse proxy is in front of a server. Intermediary so that clients never talk directly to servers.

An API gateway is a specialized type of reverse proxy. The easiest way to differentiate them is by their core focus: a load balancer focuses on scale, a reverse proxy focuses on…

RPC tends to be more for internal service communication. REST has a bit more overhead with http. Rpc tends to use more efficient formats like protobuf and grpc rather than http. This…

Either each service triggers the next one, or there is an orchestrator managing the whole thing. Queen was an orchestrator whereas in PIMS, it was choreography style, retriever triggers transformer, etc.

A service is scalable if it results in increased performance in a manner proportional to resources added. Generally, increasing performance means serving more units of work, but it can also be to…

Postgres is fine to begin with for basic search. Select where like term. When would somethint else be more relevant ? Wildcard search are slow tho, it'll scan the whole db. But…

Sticky session means it'll only work if the user has time come back to a specific server. Stateless means server needs to treat the request as if the user came in here…

Hello interview in a hurry is the best resource: https://www.hellointerview.com/learn/system-design/in-a-hurry/introduction. Book - Designing Data-Intensive Applications (haven't extracted all notes yet)

Allocate specific time slots: 5m for requirements, 20m for high-level design, 15m for deep dive, and 5m for wrapping up. Define requirements and use cases. functional requirements first, and non-functional requirements if…

https://github.com/donnemartin/system-design-primer. They have Anki flashcards too.

Throughput is the number of such actions or results per unit of time. Generally, you should aim for maximal throughput with acceptable latency.

One way to do AI Multi Agent could be to route the request to the right model depending on the complexity needed, at least to save costs. https://blog.bytebytego.com/p/how-smart-model-routing-can-cut-llm.

Reviewing code is a skill that won't go away. No shortcut to it, get good the old school way. Take the time to write the project description and code structure yourself, it…

Frontier models are good by themselves (ie Claude Code) Single agent (could) need more human in the loop to review, where multi agent could review itself.

Smolagents vs langgraph. To test the output of your models, you can rely on both deterministic and non-deterministic guardrails (LLM as Judge). Remember that AI outcome is not deterministic so it's safer…

They have an architect, scout, coder, reviewer and verifier. https://github.com/jarod-rosenthal/workermill. They have a planner, a developer (frontend, backend, devops), a tech lead. The tech lead reviews the code. Up to 3 times.

Models are so good now that they're not to blame for code quality. If there is an issue, it means there's a gap in the system/prompt/workflow.

make adr folder. make agents.md and tell claude.md to read that to be ai agnostic. write preferences to agents.md, ie testing pure functions or disallow running certain commands. but keep it light…

Good foundation is still key. AI is a signal amplifier, not just a tool. It will amplify the good and the bad. System design is still very important (some say that hyper-generalists…

We still think in black boxes and inputs/outputs like before. Collaboration is still key (in theory). We still think in processes rather than language syntax. Systems thinking. Refactoring. Engineering with ai is…

Checking if candidates memorise things is not super interesting, interviewers are interested in decision making and judgement. Communication is more important than ever.

Treat AI as a slow and non-deterministic (AI outcome is not deterministic) worker that can get things wrong, don't use it for critical paths like auth. A good chef that delegates is…

Code is cheap now. Good manually crafted code is harder to stand out in the flood of AI slop. But now everyone can generate lots of code.

Technical debt is visible but comprehension debt is not. Don't ship it if you don't understand it. Code is fast, but slow down to keep a good mental model if necessary.

If junior developers don't experiment and learn from getting things wrong, they won't develop judgement and become senior. It's important that AI doesn't replace your thinking. Juniors feel more empowered, 50% velocity…

Specialists are less required in the AI era.

In Youtube - Yannick Dalbin – Ce que 1 an à laisser l'IA coder m'a appris sur le métier de développeur, he says that if there's an issue with what the model…

useful for production apps that leverage llms beyond calling openai/claude with an api and show the results. this is useful if you need things like subagents doing different things (searching, etc.), orchestration,…

Use AI for code reviews when the risk is low.

junior people = agent with few parameters. See Youtube - EO – From Writing Code to Managing Agents. Most Engineers Aren't Ready Stanford University, Mihail Eric: context switching: switching from one task…

A model has to: find every one of those. realise they're related. ignore 99% of everything else. remember them all. compare them. That's basically Retrieval-Augmented Generation (RAG), except your retrieval system is…

Some jobs even don't say what tech stack they use, as that's not the skill they care about.

AI slop that looks ok is everywhere. In How to evaluate good engineers in the AI era, what used to be good proxies for good engineers is now different. What used to…

The market value of being an expert verificator goes up. Knowing what good looks like (having good judgement) Communication, clear intent (especially written) Context switching. A lot of these are parallels to…

Rather than modelling your data model based on what it is (relations, etc.). Model based on how the application will need to query it (key value, top 50, complex search, etc.) Optimise…

Separates compute from storage. Leverages s3 as elastic storage, making it scale way beyond what traditional dbs can do. For example, for PostgreSQL it starts struggling around 100GB to 1TB. With aurora,…

Cassandra is masterless. Cassandra has no Database ACID support. Bias towards Availability. Write heavy. Zero downtime needed (in traditional master/slave, there is a small downtime if master is down for the failover…

Removes the need for a slow sequential scan. B-tree is the most popular, allows range queries with almost O(1). Hash index (key mapping) is definitely fast and O(1) but can't do range…

Synchronous is slower, but ensures stronger consistency.

Write-heavy: Too much data for 1 instance (storage or write per second) Read-heavy: when read replicas or caching are not enough. Global queries will be slower. Complexity to operate and keep data…

2 phase commit and saga patterns are the most common ways to solve this. You can have a coordinator plan across services and commit all at once, but this is fragile if…

Amazon DynamoDB and Amazon DocumentDB are both fully managed NoSQL databases on AWS, but they are built for entirely different data models and architectural needs. Choose DynamoDB if you need predictable, single-digit…

It builds an inverted index to make searching documents by terms quickly. Example: taylor: (event1, event2) Tokenization in Elasticsearch is the process of breaking a stream of text characters down into smaller,…

1KB ish. Some common fields: But what takes most space is database overhead, index, page, etc.

PostgreSQL. Strengths: transactions, joins, constraints, flexible queries. Weaknesses: horizontal scale/sharding complexity, not ideal for huge analytical scans. DynamoDB. Strengths: massive scale, predictable low-latency key access, operational simplicity. Weaknesses: access-pattern-driven modelling, limited relational…

An image or a song would be immutable data. The access pattern for those is a lot of reads, very few writes. These would be best stored in a system where it's…

read: 20k queries per second. write: 10k writes per second.

When I'd use DDB: if key value access is enough. I need heavy write (more than 10k writes per second) automatic and managed sharding. What you're giving up with postgres: relations and…

MySQL is thread per connection, where PostgreSQL is process per connection. MySQL is seemingly faster for lots of concurrent simple reads. PostgreSQL takes up more memory, but handles complex queries better with…

Redis is great for data that needs to be accessed by key quickly. Especially if that data is temporary with the TTL feature, like a cache a ticket locking system. It would…

Memcached is supposedly a bit faster as it is multi threaded. Redis has better support of more data types. Redis has optional disk persistence.

It usually builds an inverted index for fast querying, like: Tokenization: Tokenization is the process of breaking a piece of text into individual words. This allows you to map from words to…

For example between search and main db. When introducing another database, they need to be synced somehow. Easiest: write to both in app code (what if it fails and also makes it…

Storm of: increase of people who graduated in CS, even just bootcamps. massive hires then layoffs in 2022, leaving lots of senior engineers looking for jobs. AI disrupting and making people question…

Related: + Interviewing. From: That hardest part is not building the thing, it's selling the thing.

The priority was to ensure engineers are looked after, and consulted for decisions. Related: Programming is fun. Because it was so fun, it brought a new wave of engineers. The priority was…

These things are relevant regardless of whether it's EM, TL or Staff. Or even senior eng? It gets harder to prioritise. Work on what matters. Funny enough, snacking is bad for physical…

Do not pick work that will block the team (don't become a bottleneck, be an enabler) Related: Player - Coach. Trust built through proximity to the work is different from trust built…

Is it basically doing Role - Software Engineer in the real business world?

What is engineer. Why not just developer or programmer. from What I like about software engineering, system orchestration. Related: 2 types of engineer - where they get satisfaction from.

Ic is doing what someone else wants. (Role - Software Engineer) Lead is convincing others to do what you want but never do it. Faire pour toi ça reste mieux () A…

Even in a classic 9-5 job, you're selling your services. Doing the interviews is the hard part (selling yourself). From: The software is free you are the expensive part. Related: Its harder…

https://www.youtube.com/watch?v=OVmdMcbvWmY. Related: That hardest part is not building the thing, it's selling the thing.

Chiwie. You just ship what other people want you to ship. At team level, it's what your team lead wants and the way they want it. At team lead level, it's what…

Context Action Results Learnings.

make adr folder. make agents.md and tell claude.md to read that to be ai agnostic. write preferences to agents.md, ie testing pure functions or disallow running certain commands. but keep it light…

Be comfortable with: dictionaries / sets / lists. comprehensions. iterators/generators. collections. sorting. exceptions. context managers. dataclasses. typing. JSON. HTTP. files. async basics. or. descending: Remember the distinction: creates the entire list. Whereas:

Checking if candidates memorise things is not super interesting, interviewers are interested in decision making and judgement. Communication is more important than ever.

Head of Engineering: What will be the biggest technical or organisational challenge the engineering team will face in the next year? How will this role help with that. Architecture changes? How do…

Can they do the job? Not necessarily right, now but you need to know that they can learn how to do it. Unless you have constraints that they need to be performant…

https://www.techinterviewhandbook.org/grind75/.

Allocate specific time slots: 5m for requirements, 20m for high-level design, 15m for deep dive, and 5m for wrapping up. Define requirements and use cases. functional requirements first, and non-functional requirements if…

Reminds me of That hardest part is not building the thing, it's selling the thing.

juniors execute, seniors question.

Title is misleading, there is no "should". The best managers are not removed from the craft for too long.

Or was it salary calibration. From: Programming is sometimes as much art as science.

junior people = agent with few parameters. See Youtube - EO – From Writing Code to Managing Agents. Most Engineers Aren't Ready Stanford University, Mihail Eric: context switching: switching from one task…

Whether it’s a decision about if you’ll be hired, promoted or fired. same blog: Who are you displaying your skills to? Who gets you access to opportunities? Skill alone doesn't matter. If…

If junior developers don't experiment and learn from getting things wrong, they won't develop judgement and become senior. It's important that AI doesn't replace your thinking. Juniors feel more empowered, 50% velocity…

Specialists are less required in the AI era.

The priority was to ensure engineers are looked after, and consulted for decisions. Related: Programming is fun. Because it was so fun, it brought a new wave of engineers. The priority was…

Some jobs even don't say what tech stack they use, as that's not the skill they care about.

Good foundation is still key. AI is a signal amplifier, not just a tool. It will amplify the good and the bad. System design is still very important (some say that hyper-generalists…

Do not pick work that will block the team (don't become a bottleneck, be an enabler) Related: Player - Coach. Trust built through proximity to the work is different from trust built…

What is engineer. Why not just developer or programmer. from What I like about software engineering, system orchestration. Related: 2 types of engineer - where they get satisfaction from.

Context Action Results Learnings.

Reminds me of That hardest part is not building the thing, it's selling the thing.

aka Sponsoring.

What did I build? Why did I structure it this way? Why this database? What assumptions did I make? What are the failure cases? What did I test? What doesn't it handle…

Scope (scale of the work) Ownership (initiative and driving outcome, sometimes outside your scope, SLI/SLO, incident handling) Ambiguity (making decisions with incomplete info, turning unclear goals into clear requirements, complex project) Perseverance…

Frontier models are good by themselves (ie Claude Code) Single agent (could) need more human in the loop to review, where multi agent could review itself.

Maybe not straight user ID in api. Jwt json web token contains signed json data about the user, like what role they have. Every request contains a signature that the server can…

This is when banks loan money with almost 0% interest. That usually happens after a recession to get business flowing, ie after 2008. During this period, investors take more risks to invest…

This is when banks loan money with almost 0% interest. That usually happens after a recession to get business flowing, ie after 2008. During this period, investors take more risks to invest…

It's not when it's deployed, it's when it's acknowledged by leadership and socially celebrated. If you ship something users hate and makes no money, but your leadership team is happy, you still…

Ability to demonstrate trade-offs. This is especially relevant in the AI era. Interviewing is hard: Its harder to prove someone that you can do something, than just doing the thing everyday.

WebSockets is 2 way open connection. SSE (Server Sent Events) is one way. API usually have a timeout of 30-60 seconds, configured by an outside system, often. If the operation you're doing…

Be comfortable with: dictionaries / sets / lists. comprehensions. iterators/generators. collections. sorting. exceptions. context managers. dataclasses. typing. JSON. HTTP. files. async basics. or. descending: Remember the distinction: creates the entire list. Whereas: