Haoyi's Programming Blog

Table of Contents

How to conduct a good Programming Interview

Posted 2017-08-30
Scala Vector operations aren't "Effectively Constant" timeuTest: the Essential Test Framework for Scala

Any software engineer who has ever looked for a job has had the interview experience: being locked in a small room for an hour, asked to write code to solve some arbitrary programming task, while being grilled by one or two interviewers whether or why the code they've written is correct.

You can find lots of material online about how to ace the interview as an interviewee, but very little has been written about how to conduct one as the interviewer. This post will lay out a coherent set of principles and cover many concrete steps you can use to make the most out of the next time you find yourself interviewing a potential software-engineering hire.

About the Author: Haoyi is a software engineer, an early contributor to Scala.js, and the author of many open-source Scala tools such as the Ammonite REPL and FastParse.

If you've enjoyed this blog, or enjoyed using Haoyi's other open source libraries, please chip in (or get your Company to chip in!) via Patreon so he can continue his open-source work

When recruiting software engineers, you typically conduct a whole battery of interviews to try and decide if a candidate is someone you would want to work with. While the line-up of interviews may change, there's almost always a programming interview: writing code on a whiteboard/computer, to solve some arbitrary problem, to answer the question "can this person actually write working code?".

In my work, I have conducted over a hundred programming interviews of all sorts, to candidates of all backgrounds: from new-graduates fresh out of school (or even those still in school!) to experienced engineers with decades more work experience than I have. I have driven changes in my org's interview process, such as introducing hands-on coding interviews where the candidate has to execute and debug their code on a laptop rather than just waving their hands at a whiteboard.

I have also been interviewed by more than 20 different companies over the years in the course of looking for jobs, for positions ranging from "intern" to "principal engineer". I have seen questions ranging from "fix this bug in an obscure open-source library" to "are the words desert and dessert synonyms, antonyms, or unrelated?"

It's easy to be thrown in a room and asked to interview a candidate. It's not so easy to conduct a good interview - getting a good read on a candidate's abilities and weaknesses - and to be able to do so repeatedly and reliably. This post will cover some principles and techniques you can use to make the most of the limited time you have together with a candidate, so you can make a firm decision whether or not to give the candidate a job offer to work with you in your company or on your team.


While there are a variety of ways you can conduct a programming interview, the goals typically are the same:

There are a lot of more tactical goals that you can try to accomplish, along with specific skills you may want to test (e.g. knowledge of a specific platform or programming language) but at a high level these are the primary questions you are trying to answer in a programming interview. There are other interviews (e.g. resume review, architecture design, etc.) that cover other things you'd want to know before hiring the fellow.

While you usually give the candidate a programming task to work on during the interview, it is the process through which a candidate finds a solution that reveals the most about whether they can write code. As an interviewer, you should end up with a much more holistic view of the candidate than "finished task X in Y minutes":

From these observations, you should be able to come up with answers to the four questions posed above, and then decide whether you'd call this candidate a "hire" or "no hire"

Tactical Goals

In order to achieve those high-level goals, there are some intermediate targets in the design of an interview that help you get there:


We want to test the candidate's ability to perform in something similar to the environment they'd be working in. The challenges of the interview should be challenges that are likely to appear in their day-to-day work environment.

Things like "refactoring code on a whiteboard, without the ability to insert text or copy-paste" are exceedingly challenging, but don't appear in real work. Neither do things like "remembering the API of the heapq module without looking up documentation".

Being able to refactor code they've already written to twist it in a new direction, or debugging code that is currently spitting out stack traces, is a core skill for any software engineer that very often programming interviews fail to test for.

Not everything that is challenging is realistic, and it is easy for challenging-but-silly requirements to slip in accidentally. You need to make a conscious effort to make sure they do not become the focus of the programming interview, while simultaneously spending some time focusing on the skills that are most relevant in their day-to-day work should they get hired.

Set them up for Success

You want the programming task given to be challenging, but everything else should be lined up to help the candidate succeed.

The candidate should not need to deal with vague instructions, unclear expectations, or the mind-games of trying to guess "what does the interviewer actually want". They shouldn't have to deal with antagonistic or hostile personalities.

If the real work environment is a good place to work, the interview environment should be too.

You want the candidate to be able to do their best. If you turn down a candidate, you want to be 100% sure you turned them down because, despite all efforts, they weren't able to perform. You don't want to end up hesitating, unsure whether a candidate did badly only because the interviewer was a jerk or the task given was vague and unclear.

Lastly, you want every candidate to have a good experience. Even if you don't want to hire them, they probably have friends you would want to hire! Word spreads, and remember that during an interview, the candidate is interviewing your company as well.

Maximize candidate-work time

You want to maximize the time you can watch the candidate working: writing code, tracking down bugs, discussing designs. This is your chance to judge their abilities and decide if they are capable or not.

A lot of other things often happen in interviews: you explain the task to them and make sure they understand it. You interrupt the candidate to correct them if/when they veer off course.

These are things that are necessary, but they are not the point of the interview, so you should keep them to the minimum necessary. Discussion is great, but time spent elaborating on vague & poorly delivered instructions doesn't really count as "discussion", and neither does repeated, annoying interruptions.

The time saved by minimizing these things can be given to the candidate so they has a chance to show you their stuff.

Limitations of the Programming Interview

Now that we've laid out some of the high-level and tactical goals for why we are conducting these programming interviews, it's worth spending a moment to consider the constraints of the interviews you will be conducting, and some explicit non-goals: things that we are not trying to achieve.


Random Chance

There will always be randomness in the programming interview, but it is possible to control and mitigate it.

Let's imagine you give the candidate a problem to solve, and they finish it relatively quickly and without difficulty. That could mean any of the following things:

It is impossible to pick a set of problems that every candidate is equally unfamiliar with. And if a candidate finds a solution quickly, it is very difficult to distinguish the various cases above.

However, there are things you can do that make it less random:

Through things like this, you can actively work to mitigate the effect of random chance in your interviews, even if you cannot entirely eliminate it.

Time constraints

Programming interviews tend to be anywhere from 40 minutes to an hour long. This is unlikely to change.

I have heard everyone wish they had more time in the interview process: interviewees sometimes wish they had a longer/more-realistic time frame to showcase their skills, while interviewers often lament the short and unrealistic programming tasks they are forced to give candidates to fit within the allotted time slots.

Perhaps in an "ideal interview process", the candidate would be able to spend more time with engineers from the company they are interviewing with. In practice, this quickly bumps up against reality:

There are other things you can do to try and get signal:

In the end, I think there's no silver bullet to getting a good read on a candidate. Time is limited, and we have to make the best of it to learn as much about the candidate as we can.


Successfully implementing a specific algorithm

While most often you give the candidate some kind of algorithmic task to work through, the goal for you-the-interviewer is not to try to get the candidate to finish and correctly solve the given problem!

The programming interview is like a treadmill: the point isn't to go anywhere in particular. The point is just to keep the candidate running so you can watch and analyze their technique. Due to random chance, how long a candidate takes to complete an implementation often has no bearing on how good they are.

Giving the candidate a task to solve is purely so you have a chance to watch the candidate work: it is watching them work that you can then make a decision whether or not this is a person who can write code and effectively reason about and discuss problems and constraints.

While stronger candidates do tend to on-average be able to accomplish more of the tasks that you set before them, and do so quicker, I nevertheless do not think trying to get a candidate to complete a task correctly is a goal that is worth striving towards.

Making life difficult for the candidate

It is common, not just in programming interviews, to make life difficult for the candidate. I think this idea is rubbish.

Things like being vague in what the candidate is meant to do, interrupting them while they are working, and generally being confrontational. I have seen this justified under the idea that a candidate "needs to be able to work under pressure", and thus any additional pressures you can give during an interview setting are, if not good at least acceptable.

The programming interview isn't a military boot-camp where you stress people to make the weak-willed candidates drop out; your goal is to try and get a glimpse of how the candidate thinks and works through programming problems in a realistic setting.

Hopefully in your realistic setting, the challenge comes from the difficult tasks you are given to solve, not some jerk that's sniping at you from the sidelines. You want to find candidates who would do well in the setting you'll be working with them in, not candidates who do well under hazing.

Making life difficult for the candidate, part two

On a related note, it is not the sole purpose of the interview to pose extremely difficult or challenging problems, even in the absence of unrelated difficulties. While this is something you often end up doing, remember that your goal is to get a good read of the candidate's abilities, and the difficulty of the given task is only useful so far as it helps achieve that goal.

There are multiple ways difficult problems can go against getting a good read on a candidate's abilities:

One good example I have found of the second problem is interviews which make a candidate dive into an existing large codebase and make some changes.

These interviews are extremely realistic, since diving into large codebases is what you do day-in and day-out as a software engineer, and they are also very challenging. But I have personally found that with such tasks, it is often essentially random how long it takes a candidate to find what they need. That does not give you a good signal on whether the candidate is any good.


A programming interview is not the place to nit-pick over minor details. This is simply due to the format: with less than 60 minutes to work on a task you have never seen before, often in an environment without editors or tools the candidate is comfortable with.

In a real work setting, they would have days or weeks to get comfortable with the programming environment, coding and style conventions. They will have lint rules, auto-formatters, code-review, a compiler, and many other things to make sure they doesn't forget the small stuff.

During a programming interview, none of those things are available.

Whether or they can keep track of minutiae in an interview reflects not-at-all on whether they can keep things neat in day-to-day work. Thus you should let small things slide:

During day-to-day work, you should be as strict about these things as possible: standardizing small details like this is core to the long-term health of a codebase. But during an interview, in an environment without proper tools, trying to be strict accomplishes nothing.

Before the interview

The work needed to conduct a good programming interview starts long before the candidate steps into the room. Here are some things to keep in mind when you are picking and preparing a programming task which you can then use to stretch and challenge interview candidates for months or years to come.

Pick messy, open-ended tasks

It is very common for interviewers to give programming challenges which are based on neat, elegant little algorithms, data-structures or techniques.

These elegant problems make for poor interview questions.

I have found that some of the best problems are messy and without a single, provably-optimal solution.

The problem with many tasks such as that have a single, specific solution is that it is often random how easy it is for a candidate to find that one solution. A candidate who is able to regurgitate a flawless Merge-sort implementation from memory doesn't tell you anything about if they are any good!

Another shortcoming of these problems is that there isn't much freedom designing a solution: one "correct" solution looks the same as any other: after all, how many ways can you implement a Trie? You thus miss out on seeing how the candidate writes code when they have the freedom to design things and make choices.

Specific algorithms make great practice problems because you can write code and validate it for correctness yourself without a real interviewer sitting there, but that doesn't matter at all for someone conducting an interview.

Open-ended questions, on the other hand, often have multiple solutions, and a candidate doesn't need to have studied any specific topic to put together a good solution. For example:

In these cases, it matters a lot less whether you are familiar with any specific algorithm or data-structure. The techniques you need to solve these problems are a lot more ubiquitous, and you have a lot more different ways of approaching the problem. This means:

Note that this doesn't make the problem any less difficult! There are still plenty of opportunities for a candidate to screw up, to get tangled up in their own code, or to demonstrate their fluency and competence. Even if the task proves to be too simple, such problems can easily be extended into multi-stage tasks to keep things challenging.

What's important is that open-ended sorts of tasks try and minimize the advantage/disadvantage a candidate will get from mostly random chance decisions: knowing a specific algorithm or arbitrarily picking a particular approach. Rather than having a single solution that the candidate either gets or doesn't, such questions should allow any candidate to start working and iterating towards a good solution, giving the interviewer lots of time to see them work.

Prepare multi-stage tasks

You should design your tasks such that they are broken down into multiple related stages.

Earlier, I mentioned that it is a non-goal to get the candidate to successfully implement a specific algorithm. Taken to the extreme, ideally you should be able to keep a candidate working hard for the full duration of the interview, no matter how good they are. Obviously, if you have a single task some candidates will finish it earlier and some later. The solution, then, is to have a multi-stage task, comprised of many steps each dependent on another.

For example, a simple task might be:

Implement an arithmetic expression parser that can evaluate strings like (1+2)*(3+(4*5)) to 69. Assume every binary operation a+b must be wrapped in a set of parens (a+b)

That may take someone anywhere from 10 minutes to 30 minutes to complete. Maybe someone who has seen similar problems before will be able to directly spit out a working solution in 5 minutes. What then?

If someone finishes it early, you should have follow-on task after follow-on task to keep them working hard, so you can watch them continue working.


Allow multiple operations within a single set of parens, or multiple sets of wrapping parens, allowing for operator precedence (* and / are higher precedence than + and -) e.g. evaluating ((1+2))*(3+4*5) into 69


Extend it to allow variables in the expression, e.g. evaluating (y+2)*(3+x*5) into a object f, and later call f(1,4) to return 69

And then:

Limiting the only operations to + and *, extend your code to expand the symbolic expression, e.g. converting (y+2)*(3+x*5) into 5yx + 3y + 10x + 6

Remember, the programming interview is like a treadmill: the point isn't to go anywhere in particular. The point is just to keep the candidate running.

Having a sequence of multiple related tasks like this helps you mitigate the effect of random chance: even if they've seen your initial task before, they're unlikely to have seen all the follow-ups.

Compared to having multiple unrelated tasks, or a single task with all requirements provided up-front, having multiple related tasks with additional requirements intentionally not provided up front forces the candidate to repeatedly refactor their old code to make it do a new thing. This is usually a fascinating process, and teaches you more about how a candidate works than how fast they can spit out the code to reverse a binary tree.

Lastly, having multi-stage tasks results in having lots of natural stopping points: one after each stage. Those are natural points for you to interject for discussion, or to wrap up the interview seamlessly if time is running short.

Test your programming task

Whatever task you end up giving to a candidate, should first be tested on one, or preferably more than one, of your colleagues.

By "tested", I mean that you should sit down with your colleague(s) in a room, get them to do the interview task under interview time-constraints, interacting with them as you would in an interview setting. Whatever solution they came up with, how it long it took for them to complete various stages of the problem, and any other notable facts should all be written down and tabulated in a document somewhere that any interviewer administering this question will have access to.

What does testing the interview task before-hand give you?

Testing an interview task is somewhat tedious and annoying: rather than simply asking the question when a candidate comes in, you now have to ahead-of-time budget 1-2 hours with your colleagues to administer the question to them and discuss/tabulate the results after.

Nevertheless, you will probably conduct many interviews over the course of your career, and will no-doubt re-use the same interview questions many times. Thus this relatively-small up-front investment is definitely worth it for the many ways it helps you make the most out of the your limited time with a candidate in the interview room.

Starting the Interview

You've done all your prep work. You're now waiting in the interview room, and someone has noticed the candidate hanging around the front door and brought them in.

Even now, as the interview is just starting, there are things you can screw up. Here are some things to keep in mind during the first few minutes of your hour-long slot:

Map out the structure of the interview before starting

After introductions, the first thing that you should do is go over the structure of the interview.

One common layout I've seen is for 60-minute interview slots is:

This can obviously vary with the overall length of the interview and personal preference. Regardless of what layout you choose, you should state it explicitly. This helps the candidate:

Overall, it helps set them up for success, since the candidate can feel secure and produce their best work, and is definitely worth the 15 seconds or so it takes to deliver to them.

Clearly separate what is expected from what is provided

You need to make sure you are clear what you are meant to provide the candidate, and what the candidate is expected to produce in return.

In many problems you'll need to provide back-story, constraints, requirements, or stub functions and APIs for the candidate to make use of. While sometimes you can get away with asking the candidate e.g. "what functions do you think you'll need" or "what do you think the requirements would be", taken too far this just results in a confused/frustrated candidate and a confused/frustrated interviewer unable to make progress.

Conversations like:

Candidate: This solution takes O(log n) time per operation

Interviewer: Do you think that's fast enough?

Candidate: Yeah it seems fast enough for me

Interviewer: Would you be happy deploying this to production?

Candidate: Since you haven't told me how important production is, sure why not

In real work, there will be external context and limitations on the system, timelines and budgets, and stakeholders who will have requirements that need to be met. In an interview settings, none of that exists, so it is up to the interviewer to provide it.

A candidate trying to propose a solution without being provided the necessary constraints and requirements is like an engineer trying to build a system without knowledge of the system's requirements or access to stakeholders: a frustrating, useless waste of time that's sure to end in disaster.

If you ask the candidate the define the requirements for the task you're going to give them, you can't be surprised if the candidate imagines a use case with entirely different requirements from what you expected!

While it's good to have discussions with the candidate to learn how they think, you as-the-interviewer need to have a crystal-clear understanding of what it is you want the candidate to deliver and what it is you are willing to provide. No matter how capable a candidate is, they cannot deliver something if you don't provide the context/constraints/requirements necessary.

Once that is clear, you will always be able to drive the discussion towards a clear destination: either you providing something to the candidate, or the candidate delivering something to you. This clarity, for both the interviewer and the candidate, goes a long way to keeping the interview flowing smoothly and allowing focus on the task at hand.

Provide the task description printed out on a piece of paper

Before an interview, you should just type out the problem statement, and hand them the piece of paper. This should contain everything they need: any background info, the problem description, example inputs and outputs. Something like:

Your task, should you choose to accept it, is to implement a LRU cache.

LRU caches are often used to implement caches which you do not want to grow indefinitely. The cache has a max size, and when a new key is inserted that would make it grow larger than the max size, the key which has not been accessed for the longest time is removed to make space.

It should support these operations:

What happens if you call get(key) on a key which doesn't exist in the cache is up to you to define.

Example usage:

cache = new LruCache(2) // capacity = 2

cache.put(1, "1")
cache.put(2, "2")
cache.get(1)          // returns "1"
cache.put(3, "3")     // evicts key 2, because the key 1 was retrieved
cache.get(2)          // returns null, because 2 was just evicted
cache.put(4, "4")     // evicts key 1,
cache.get(1)          // returns null, because it was evicted earlier
cache.get(3)          // returns "3"
cache.get(4)          // returns "4"

If there are any questions, then the interviewer can answer them.

This accomplishes a lot of things:

The explanation of the problem statement at the start of the interview is basically boilerplate: tedious, repetitive, but very important, and easy to get wrong and cause frustration or confusion. "How well a candidate clarifies poorly stated requirements" is certainly something you could test for, but isn't generally one of of the main goals of the programming interview.

A standard printout ensures a basic level of consistency between candidates (they all got the exact same problem), and allows the candidate to quickly understand the basic problem they need to solve so they can move on to higher-level discussion and implementation. It is that discussion and implementation that is the real meat of the programming interview.

During the Interview

The interview is now well under way: you have 45 minutes to try and figure out if this candidate is someone you are willing to hire and work with on your team.

Here are some tips to optimize the main body of the interview, to try and maximize the signal you get out of your limited time slot so you can be clear and confident in your final yes/no judgement.

Write code on a computer, not a whiteboard

As an interviewer you should default to getting the candidate to type out their solution, whether they want to or not. Many programming interviews still take place on a whiteboard. As a candidate, if you ask, they'll often let you type it out on a computer, but this is inadequate: the computer should be the default.

Writing code on a whiteboard always ends in a mess that confuses everybody, benefits nobody and is an pain in the neck to read. There are many strange things that happen when you start writing code on a whiteboard:

While these things take up time, none of them are relevant at all to what the candidate will be expected to do day-to-day if they get hired. Nobody spends time writing and refactoring large blobs of code on a whiteboard in their day to day work.

To maximize realism and maximize candidate-work time, you should make the candidate write code on a computer with all the modern conveniences of "insert text at line" and "copy-paste", or even things like syntax highlighting. While it's obviously ideal if the candidate can write code on their own machine in an environment they're familiar with, any computer with a text editor is already a huge step up in efficiency, clarity and realism.

You can keep a whiteboard around, for drawing diagrams and such that text editors do not do well. They just shouldn't be the default place to write, edit and discuss chunks of code.

Make the candidate run and debug their code

You should, if possible, make the candidate execute and debug the code they write on a computer.

Of the places I've interviewed with, only a small minority of them by-default expected you to execute-and-debug the code you wrote; the vast majority of interviewers just wanted to inspect the code themselves for correctness. I think this is the wrong default.

Most people who have done a programming interview (either as interviewer or interviewee) would have an experience like:

Interviewer: I see a bug in your code

Candidate: It looks correct to me

Interviewer: Are you sure?

Candidate: Yes

Interviewer: What if you put in this input?

Candidate: It looks like it does the right thing

Interviewer: The bug I see is in this function f. Do you see it?

Candidate: No, it looks correct

This back-and-forth can take forever. Sometimes the interviewer is right, sometimes the candidate is right. Always though, this kind of exchange is a waste of time, and only occurs because you're not executing the code being written.

If you actually run the code, the exchange could instead go:

Interviewer: I see a bug in your code

Candidate: It looks correct to me

Interviewer: Are you sure?

Candidate: Yes

Interviewer: What if you put in this input?

Computer: IndexError: index out of range: 1, File "main.py", line 1, in <module>

And thus the candidate is given irrefutable proof that something is broken, without needing to explain to them what the problem is. They can then use their standard debugging techniques to figure out what is wrong, and fix it.

While not every programming exercise can be conveniently packaged up in an executable form (e.g. some involve calling mock functions or APIs that don't actually exist), many can. When you're asking someone to serialize a binary tree to a linked list, there's no reason why the candidate should not be able to take their code, run it, and either verify its correctness or begin the debugging process.

Making the candidate execute-and-debug their code is a very different interview experience than just having the interviewer inspect it. I think it has a number of benefits, some of them quite surprising:

Overall, writing-and-executing code is a mindset change compared to writing code and having the interviewer squint at it. My experience is that those interviews go much smoother, are much less stressful for both parties, candidates write better code and nevertheless the interviewer ends up with much better signal on how fluent the candidate is at writing code to solve problems.

Let the candidate make mistakes

It is ok to let a candidate explore approaches or solutions you know will end up not working.

Very often, I've seen a candidate start ever-so-slightly down the "wrong" path to the solution, only to be immediately and vigorous stopped by the interviewer. They are then forced, via constant questioning, back onto the "correct" path, until they later diverge and again have to be stopped.

This kind of interviewer behavior isn't entirely unexpected: as engineers, we loathe to see someone doing things the "wrong way", and want to help them get to the correct answer as quickly as possible. This is something that is of value in our day-to-day work.

But interviews are not like day-to-day work.

One of the main ways interviews are not like day-to-day work is that it is a non-goal to get the candidate to successfully implement a specific algorithm.

The point of the interview task is to make the candidate work. It is almost irrelevant whether or not the task ends up being completed.

Thus, you should feel free to let a candidate make the "wrong" decisions.

Refactoring buggy code to turn it into correct code is one of the things that software engineers do day-in and day-out. By letting the candidate run with a mistake for a while, you then get a chance to see how well they do at this very crucial skill. Watching how someone rescues a broken piece of code they wrote tells you much more than watching them regurgitate a straightforward breadth-first-search for the n-th time.

Remember that getting the candidate to the "correct" solution is not what your job as the interviewer is. Your job is to see the candidate work in a variety of realistic scenarios, and from the way they work judge how fluent they are at the work they'll have to do if you end up hiring them. Working with-broken-code/under-broken-assumptions for a while and having to debug/refactor/fix/recover after is definitely an all-too-realistic scenario for people working in this industry.

Minimize Interruptions

You should consciously minimize the number of times you interrupt a candidate while they are working.

Remember, the goal is not to get the candidate to successfully implement a specific algorithm, and that it is ok to let the candidate make mistakes. If a candidate spending some time on an (eventually) dead-end approach gives you a chance to watch them think and work, that is no loss at all given the goals of the programming interview.

Minimizing interrupts helps avoid making life difficulty for the candidate, since interrupts are inherently an uncomfortable, confrontational affair. Furthermore, by interrupting you are reducing the amount of time you can watch the candidate work, which you want to maximize. While you want to know if the candidate can discuss code, being interrupted repeatedly in the middle of work is something most software engineers would agree is not a welcome form of "discussion".

It's OK to interrupt occasionally, but you can often be strategic so your interrupts are as un-disruptive as possible:

Interrupting the candidate in the middle of work, while sometimes unavoidable, is something that you can and should take care to minimize.

After the interview

After you're 60-minutes with the candidate are over, hopefully you have a pretty good sense of how they are as an engineer, and whether you'd like to work with them.

It's then time to write up your experience, both objective, e.g.

And subjective, e.g.

Ideally, each subjective point would have one-or-more objective points substantiating it. This is much harder than just writing down subjective feelings, but forces you to be slightly more rigorous about why you feel a certain way about the candidate. It also makes it much easier when you have multiple people's write-ups that need to be integrated into a coherent, holistic picture.

Depending on your process, the candidate may have multiple interviews to complete. Perhaps the interviewers will get together to exchange notes about the candidate and discuss their experience, or perhaps someone will consolidate the feedback. Eventually, someone will make the call whether or not to give the candidate a job offer. And that's the end of the interview process!


It's not hard to be thrown in a room and asked to give a programming interview. However, it is also not hard to conduct the interview poorly: antagonizing the interviewee, testing irrelevant skills, or wasting precious minutes on things which don't matter. Remember that these goals are the things that actually matter:

There are countless materials available on how to do well in a programming interview as the interviewee: books, blog posts, skills-training websites, and countless other sources. But there is almost nothing written about how to conduct a good interview as the interviewer. This post tries to fill that void.

This post has hopefully described a coherent philosophy behind the programming interview, as well as several concrete steps that you can take to try and make your interviews both smooth and insightful.

While in no way comprehensive, hopefully the guidelines presented here can help you optimize the programming interviews you do conduct, so you can maximize the amount you can learn about a software engineering candidate in the limited time you have.

Scala Vector operations aren't "Effectively Constant" timeuTest: the Essential Test Framework for Scala