4 ms·

>The hot path is worth optimizing first, in my mind -- giving up perf there for stats which will be called much less doesn't make sense to me. You aren't getti

>The hot path is worth optimizing first, in my mind -- giving up perf there for stats which will be called much less doesn't make sense to me. You aren't getting what I'm saying. logN is already a huge optimization for a hot path. Look at that pic: https://dev-to-uploads.s3.amazonaws.com/i/ya7xk4n9ch50xgwibt9t.png https://dev-to-uploads.s3.amazonaws.com/i/ya7xk4n9ch50xgwibt... Seriously look at O(logN) and look at NlogN. Databases everywhere already take logN as acceptable. You are optimizing past that and introducing an NlogN sort step that will block all hot path calls because redis is single threaded. I don't think it's a to each their own thing here, i believe the user experience metrics will definitively play out towards my methodology but we can't know until we actually have metrics so we're at an impasse here. >This assertion is so wide that it cannot be true. No matter what language you're in, you can absolutely do sorting in the API layer. No this is absolutely true. Especially if you're a nodejs programmer. Nodejs is single threaded, sorting imposes NlogN penalty on a single thread. Backend web development for the api layer is all about transmitting as much compute as possible to the database layer. You can get away with small sorts and things can still work but in general the practice of backend optimization is indeed to let the database handle as much as possible. We can debate it but from my experience I know I'm pretty much categorically right on this. It's a fundamental principle. And it has special application to things that use Async IO like NodeJS. That's why there's huge emphasis in nodejs in writing non-blocking code. A sort is a potential block. There are very few cases where you would handle compute on the api server rather then the database. The first instance is obvious, if N is guaranteed to be small which is not the case in the takehome. In other instances of sort with large N, the server ideally needs to launch that sort in a separate non-blocking thread so that the server can continue handling oncoming requests. In nodejs, to my knowledge since I last used it, you can't easily spawn a thread at will, you just have to pray you have available workers that aren't blocked by a sorting operation. In python the GIL makes this more complicated, but slightly better then nodejs. For node you probably have 8 workers on 8 CPU cores. If you have 100 requests/s, all you need is 8 of them calling a sort on large N and all threads become blocked. Don't believe me? Read the official docs of nodejs: https://nodejs.org/en/docs/guides/dont-block-the-event-loop#don't-block-the-event-loop https://nodejs.org/en/docs/guides/dont-block-the-event-loop#... NlogN stands at the cusp here. It could work for small N, but for large N it will fail. Anything above NlogN is a huge risk. >You don't need to pipeline if the operation is O(1) -- it's not a question of whether transactions are good/bad, I don't know why you're assuming people don't know about them. This is categorically false. The pipeline operation as I said HAS NOTHING to do with O(1) or O(N). This is essentially the operation: set["key"] += 1 # another thread can update the value here leading to print displaying something unexpected. print(set["key"]) So to solve the above comment you have to place both operations in a transaction. No two ways about it. The above operation is in spirit what I am doing. I update a value, then I read the value. Two operations and a possible race condition in between. >Again, this is a matter of what people think is hard and what isn't. Integration testing does not require that you spin up docker compose -- I literally linked to a libraries that do this in a reusable (importable) way w/ docker. You linked? Where? I'd like to know about any library that will do this. Tell me of any library that does integration tests that spins up infrastructure for you. The only one closest I can think of that you can run locally is anything that would use docker-compose or some other IAC language that controls containers. I honestly don't think any popular ones exist. >If docker was the problem, you can open subprocesses and assume that redis is installed in the test environment (i.e. your laptop, or CI environment), or take a binary to the redis binary path from ENV. No way I'm going to assume the user has redis installed on his local machine. Many devs don't. It's all remote development for them or everything lives in containers. Docker is not the problem. Docker or virtual machines makes this problem more amenable to a solution, but even using docker here with testing is overkill and hacky. A take home should not expect a user to build excessive infrastructure locally just to run tests. >You said you don't do this kind of testing regularly, so it is understandable that it's not easy for you. I'd like to know your perspective. How is spinning up infrastructure for integration tests easy? You have to write code that launches databases, launches servers and views the entire infra externally. Do-able? Yes? Easy? depends on what you call easy. As easy as unit tests? Definitely no. Easy enough for a take home? In my opinion, no. Maybe for this take home project you could be right. I could do some integration tests by spinning up docker-compose from within python. Hacky but doable. But in general this solution is not production scalable as production involves more things then what can be placed inside a docker-compose. >You can even remove the burden of running redis from the test suite and just make sure it's running manually or in test suite invocation/setup. There are many options, and it's just not that hard in 2023. So I should expect the reviewer to have redis running on his local machine? I would expect in 2023, I should have the entire project be segregated from the context it's running in. That's the cutting edge I believe. Yeah it's 2023, you tell me how integration testing should be done as easily as unit testing on a takehome. I had one takehome project involving S3 and aws lambdas. They expected me to get an AWS account and literally set up infrastructure because there's no way around even testing what I wrote without actual infrastructure. That entire project was just integration test nightmare. Much rather run asserts locally. >I referred to the context of the project -- you did* have full control. Well I mean 99% of hires get into a context where they aren't in full control. So why is it logical to test what a developer will do if he/she had full control? Isn't it better to test the developers ability to code and to adapt to different contexts? Your methodology sort of just tests the programmers personal philosophy and whether it matches your own. That was my point.

hardwaresofton 3y ago
> You aren't getting what I'm saying. logN is already a huge optimization for a hot path. Look at that pic: https://dev-to-uploads.s3.amazonaws.com/i/ya7xk4n9ch50xgwibt https://dev-to-uploads.s3.amazonaws.com/i/ya7xk4n9ch50xgwibt... > > Seriously look at O(logN) and look at NlogN. Databases everywhere already take logN as acceptable. You are optimizing past that and introducing an NlogN sort step that will block all hot path calls because redis is single threaded. I don't think it's a to each their own thing here, i believe the user experience metrics will definitively play out towards my methodology but we can't know until we actually have metrics so we're at an impasse here. Just to keep the comparison in line -- we're comparing O(log(N)) @ O(1) in redis-land. On the stats side we're comparing O(N) on the redis side PLUS either O(Nlog(N)) or O(N) on the python side. > No this is absolutely true. Especially if you're a nodejs programmer. Nodejs is single threaded, sorting imposes NlogN penalty on a single thread. NodeJS has both child processes and workers for heavy computation: - https://nodejs.org/api/child_process.html https://nodejs.org/api/child_process.html - https://nodejs.org/api/worker_threads.html https://nodejs.org/api/worker_threads.html If you want to parallelize work on a thread pool there are libs that do just that: - https://www.npmjs.com/package/piscina https://www.npmjs.com/package/piscina ((this is also part of the reason I think NodeJS is strictly superior to other "scripting" languages as mentioned briefly in the other comment)) Again, here you seem to be arguing against a strawman that doesn't know that blocking the IO loop is bad. Try arguing against one that knows ways to work around that. This is why I'm saying this rule isn't true. Extensive computation on single-threaded "scripting" languages is possible (and even if it wasn't, punt it off to a remote pool of workers, which could also be NodeJS!). All of this is a premature optimization, but my point here is that the stats can be optimized later. You could literally compute it periodically or continuously and only return the latest cached version w/ a timestamp. This makes the stats endpoint O(1). > This is categorically false. The pipeline operation as I said HAS NOTHING to do with O(1) or O(N). > ... I think this is where we're talking past each other, so let me explain more of how I see the problem -- the solution I have in mind is serializing the URL and using ONE call to INCR (https://redis.io/commands/incr/ https://redis.io/commands/incr/) on the ingest side. There is a lot you can do with the data storage pattern to make other operations more efficient, but on the stats side, the most basic way you can do it is to scan I will concede that given that we know the data should fit in memory (otherwise you just crash) your approach gives you O(N) retrieval time and it's definitely superior to not have to do that on the python side (and python just streaming the response through). I am comfortable optimizing in-API computation, so I don't think it's a problem. Here's what I mean -- you can actually solve the ordering problem in O(N) + O(M) time by keeping track of the max you've seen and building a sparse array and running through every single index from max to zero. It's overkill, but it's generally referred to as a counting sort: https://ebrary.net/81651/geography/sorting_algorithms https://ebrary.net/81651/geography/sorting_algorithms This is overkill, clearly, and I will concede that ZSET is lighter and easier to get right than this. > You linked? Where? I'd like to know about any library that will do this. Tell me of any library that does integration tests that spins up infrastructure for you. The only one closest I can think of that you can run locally is anything that would use docker-compose or some other IAC language that controls containers. I honestly don't think any popular ones exist. https://testcontainers-python.readthedocs.io/en/latest/ https://testcontainers-python.readthedocs.io/en/latest/ I am very sure that I linked that, but in the case I didn't, here it is again -- hope you find it useful. > No way I'm going to assume the user has redis installed on his local machine. Many devs don't. It's all remote development for them or everything lives in containers. > > Docker is not the problem. Docker or virtual machines makes this problem more amenable to a solution, but even using docker here with testing is overkill and hacky. A take home should not expect a user to build excessive infrastructure locally just to run tests. I don't think these statements make sense -- having docker installed and having redis installed are basically equivalent work. At the end of the day, the outcome is the same -- the developer is capable of running redis locally. Having redis installed on your local machine is absolutely within range for a backend developer. Also, remote development is not practiced by many companies -- the only companies I've seen doing thin-clients that are large. > Maybe for this take home project you could be right. I could do some integration tests by spinning up docker-compose from within python. Hacky but doable. But in general this solution is not production scalable as production involves more things then what can be placed inside a docker-compose. I see it as just spinning up docker, not compose -- you already have access to the app (ex. if it was buildable via a function) so you could spawn redis in a subprocess (or container) on a random port, and then spawn the app. I agree that it is not trivial, but the value is high (in mymind. > Yeah it's 2023, you tell me how integration testing should be done as easily as unit testing on a takehome. I had one takehome project involving S3 and aws lambdas. They expected me to get an AWS account and literally set up infrastructure because there's no way around even testing what I wrote without actual infrastructure. That entire project was just integration test nightmare. Much rather run asserts locally. I agree that integration testing is harder -- I think there's more value there. Also, for replicating S3, minio (https://github.com/minio/minio https://github.com/minio/minio) is a good stand-in. For replicating lambda, localstack (https://docs.localstack.cloud/user-guide/aws/lambda/ https://docs.localstack.cloud/user-guide/aws/lambda/) is probably reasonable there's also frameworks with some consideration for this (https://www.serverless.com/framework/docs/providers/aws/guide/testing https://www.serverless.com/framework/docs/providers/aws/guid...) built in. That said I do think that's a weakness of the platform compute stuff -- it is inconvenient to test lambda outside of lambda. > Well I mean 99% of hires get into a context where they aren't in full control. So why is it logical to test what a developer will do if he/she had full control? Isn't it better to test the developers ability to code and to adapt to different contexts? Your methodology sort of just tests the programmers personal philosophy and whether it matches your own. That was my point. Ah, this is true -- but I think this is what people are testing in interviews. There is a predominant culture/shared values, and the test is literally whether someone can fit into those values. Whether they should or should not be, that's at least partially what interviews are -- does the new team member feel the same way about technical culture currently shared by the team. Now in the case of this interview your solution was just fine, even excellent (because you went out of your way to do async io, use newer/easier packaging methodologies, etc), but it's clearly not just that.

hardwaresofton 3y ago

> You aren't getting what I'm saying. logN is already a huge optimization for a hot path. Look at that pic: https://dev-to-uploads.s3.amazonaws.com/i/ya7xk4n9ch50xgwibt https://dev-to-uploads.s3.amazonaws.com/i/ya7xk4n9ch50xgwibt... > > Seriously look at O(logN) and look at NlogN. Databases everywhere already take logN as acceptable. You are optimizing past that and introducing an NlogN sort step that will block all hot path calls because redis is single threaded. I don't think it's a to each their own thing here, i believe the user experience metrics will definitively play out towards my methodology but we can't know until we actually have metrics so we're at an impasse here. Just to keep the comparison in line -- we're comparing O(log(N)) @ O(1) in redis-land. On the stats side we're comparing O(N) on the redis side PLUS either O(Nlog(N)) or O(N) on the python side. > No this is absolutely true. Especially if you're a nodejs programmer. Nodejs is single threaded, sorting imposes NlogN penalty on a single thread. NodeJS has both child processes and workers for heavy computation: - https://nodejs.org/api/child_process.html https://nodejs.org/api/child_process.html - https://nodejs.org/api/worker_threads.html https://nodejs.org/api/worker_threads.html If you want to parallelize work on a thread pool there are libs that do just that: - https://www.npmjs.com/package/piscina https://www.npmjs.com/package/piscina ((this is also part of the reason I think NodeJS is strictly superior to other "scripting" languages as mentioned briefly in the other comment)) Again, here you seem to be arguing against a strawman that doesn't know that blocking the IO loop is bad. Try arguing against one that knows ways to work around that. This is why I'm saying this rule isn't true. Extensive computation on single-threaded "scripting" languages is possible (and even if it wasn't, punt it off to a remote pool of workers, which could also be NodeJS!). All of this is a premature optimization, but my point here is that the stats can be optimized later. You could literally compute it periodically or continuously and only return the latest cached version w/ a timestamp. This makes the stats endpoint O(1). > This is categorically false. The pipeline operation as I said HAS NOTHING to do with O(1) or O(N). > ... I think this is where we're talking past each other, so let me explain more of how I see the problem -- the solution I have in mind is serializing the URL and using ONE call to INCR (https://redis.io/commands/incr/ https://redis.io/commands/incr/) on the ingest side. There is a lot you can do with the data storage pattern to make other operations more efficient, but on the stats side, the most basic way you can do it is to scan I will concede that given that we know the data should fit in memory (otherwise you just crash) your approach gives you O(N) retrieval time and it's definitely superior to not have to do that on the python side (and python just streaming the response through). I am comfortable optimizing in-API computation, so I don't think it's a problem. Here's what I mean -- you can actually solve the ordering problem in O(N) + O(M) time by keeping track of the max you've seen and building a sparse array and running through every single index from max to zero. It's overkill, but it's generally referred to as a counting sort: https://ebrary.net/81651/geography/sorting_algorithms https://ebrary.net/81651/geography/sorting_algorithms This is overkill, clearly, and I will concede that ZSET is lighter and easier to get right than this. > You linked? Where? I'd like to know about any library that will do this. Tell me of any library that does integration tests that spins up infrastructure for you. The only one closest I can think of that you can run locally is anything that would use docker-compose or some other IAC language that controls containers. I honestly don't think any popular ones exist. https://testcontainers-python.readthedocs.io/en/latest/ https://testcontainers-python.readthedocs.io/en/latest/ I am very sure that I linked that, but in the case I didn't, here it is again -- hope you find it useful. > No way I'm going to assume the user has redis installed on his local machine. Many devs don't. It's all remote development for them or everything lives in containers. > > Docker is not the problem. Docker or virtual machines makes this problem more amenable to a solution, but even using docker here with testing is overkill and hacky. A take home should not expect a user to build excessive infrastructure locally just to run tests. I don't think these statements make sense -- having docker installed and having redis installed are basically equivalent work. At the end of the day, the outcome is the same -- the developer is capable of running redis locally. Having redis installed on your local machine is absolutely within range for a backend developer. Also, remote development is not practiced by many companies -- the only companies I've seen doing thin-clients that are large. > Maybe for this take home project you could be right. I could do some integration tests by spinning up docker-compose from within python. Hacky but doable. But in general this solution is not production scalable as production involves more things then what can be placed inside a docker-compose. I see it as just spinning up docker, not compose -- you already have access to the app (ex. if it was buildable via a function) so you could spawn redis in a subprocess (or container) on a random port, and then spawn the app. I agree that it is not trivial, but the value is high (in mymind. > Yeah it's 2023, you tell me how integration testing should be done as easily as unit testing on a takehome. I had one takehome project involving S3 and aws lambdas. They expected me to get an AWS account and literally set up infrastructure because there's no way around even testing what I wrote without actual infrastructure. That entire project was just integration test nightmare. Much rather run asserts locally. I agree that integration testing is harder -- I think there's more value there. Also, for replicating S3, minio (https://github.com/minio/minio https://github.com/minio/minio) is a good stand-in. For replicating lambda, localstack (https://docs.localstack.cloud/user-guide/aws/lambda/ https://docs.localstack.cloud/user-guide/aws/lambda/) is probably reasonable there's also frameworks with some consideration for this (https://www.serverless.com/framework/docs/providers/aws/guide/testing https://www.serverless.com/framework/docs/providers/aws/guid...) built in. That said I do think that's a weakness of the platform compute stuff -- it is inconvenient to test lambda outside of lambda. > Well I mean 99% of hires get into a context where they aren't in full control. So why is it logical to test what a developer will do if he/she had full control? Isn't it better to test the developers ability to code and to adapt to different contexts? Your methodology sort of just tests the programmers personal philosophy and whether it matches your own. That was my point. Ah, this is true -- but I think this is what people are testing in interviews. There is a predominant culture/shared values, and the test is literally whether someone can fit into those values. Whether they should or should not be, that's at least partially what interviews are -- does the new team member feel the same way about technical culture currently shared by the team. Now in the case of this interview your solution was just fine, even excellent (because you went out of your way to do async io, use newer/easier packaging methodologies, etc), but it's clearly not just that.