“Infinity” is a Bad Default Timeout

March 17, 2017

Many libraries wrap some external communication. Be it a REST-like API, a message queue, a database, a mail server or something else. And therefore you have to have some timeout – for connecting, for reading, writing or idling. And sadly, many libraries have their default timeouts set to “0” or “-1” which means “infinity”.

And that is a very useless and even harmful default. There isn’t a practical use case where you’d want to hang on forever waiting for a resource. And there are tons of situations where this can happen, e.g. the other end gets stuck. In the past 3 months I had 2 libraries that have a default timeout of “infinity” and that eventually lead to production problems because we’ve forgotten to configure them properly. Sometimes you even don’t see the problem, until a thread pool gets exhausted.

So, I have a request to API/library designers (as I’ve done before – against property maps and encoding other than UTF-8). Never have “infinity” as a default timeout. Your library will thus cause lots of production issues.
Also note that it’s sometimes an underlying HTTP client (or Socket) that doesn’t have a reasonable default – it’s still your job to fix that when wrapping it.

What default should you provide? Reasonable. 5 seconds maybe? You may (rightly) say you don’t want to impose an arbitrary timeout on your users. In that case I have a better proposal:

Explicitly require a timeout for building your “client” (because these libraries are most often clients for some external system). E.g. Client.create(url, credentials, timeout). And fail if no timeout is provided. That makes the users of the client actively consider what is a good timeout for their usecase – without imposing anything, and most importantly – without risking stuck connections in production. Additionally, you can still present them with a “default” option, but still making them explicitly choose it. For example:

Client client = ClientBuilder.create(url)
// OR
Client client = ClientBuilder.create(url)

The builder above should require “timeouts” to be set, and should fail if neither of the two methods was invoked. Even if you don’t provide these options, at least have a good way of specifying timeouts – some libraries require reflection to set the timeout of their underlying client.

I believe this is one of those issues that look tiny, but caus a lot of problems in the real world. And it can (and should) be solved by the library/client designers.

But since it isn’t always the case, we must make sure that timeouts are configured every time we use a 3rd party library.

Share Button

3 Responses to ““Infinity” is a Bad Default Timeout”

  1. I agree entirely! There is a lot to be said for “sensible defaults”. In the case of a socket connection, a 10 seconds timeout probably suffices. Maybe a 30 second timeout for a REST call.
    In fact, why not have a 5 second timeout and retry instead and bail out if second call fails too?
    There are many ways to provide a simple (!!) sensitive timeout and we all ought to do it.

    A techie burnt by exhausted threadpools for similar reasons 🙂

  2. FYI : https://github.com/Netflix/Hystrix

  3. Last two problems in prod:

    – “Infinity” HTTP Default Timeout
    – “Infinity” SMTP Default Timeout


Leave a Reply