Hacker News Reader

Comment by zaphar | Hacker News Reader

(full disclosure: I work at google and also like erlang)

Erlang has fantastic facilities for robustness and concurrency. What it does not have is type safety and it's terrible at handling text in a performant fashion. So if you don't care about either of those things and only care about robustness and concurrency then Erlang is great. There were internal discussions about Erlang here but the upshot was. We had already basically duplicated Erlangs supervision model in our infrastructure, only we did it for all languages and Erlang didn't offer any benefits in performance for us. It's only benefit would have been the concurrency model. That's much less benefit than Go gives.

Go gives you Erlangs concurency model, a similar philosophy of robustness, type safety, all batteries included, and performance. Equating the two languages works in 1 or 2 dimensions but not on all the dimensions google cares about.

View on HN

Interesting, thanks for that; it's pretty much what I guessed (especially the bit about the supervision tree and hot-code-upgrade advantages being mooted by your infrastructure.)

On a tangent, though:

> What it does not have is type safety

I've tried to work this out before (I'm designing a new language for Erlang's VM), but as far as I can tell, type safety is in practice incompatible with VM-supported hot code upgrade.

If you have two services, A and B, and you need to upgrade them both, but you can't "stop the world" to do an atomic upgrade of both A and B together (because you're running a distributed soft real-time system, after all), then you need to switch out A, and then switch out B.

So, at some point, on some nodes, A will be running a version with an ABI incompatible with B. In a strongly-typed system, the VM wouldn't allow A's new code to load, since it refers to functions in B with type signatures that don't exist.

On the other hand, in a system with pattern-matching and a "let it crash" philosophy, you just let A's new code start up and repeatedly try-and-fail to communicate with B for a while, until B's code gets upgraded as well--and now the types are compatible again.

It's an interesting problem.

> type safety is in practice incompatible with VM-supported hot code upgrade.

That's not true.

First, it's very easy to hot reload changes that have been made to the code that are backward compatible. The JVM spec describes in very specific details what that means (adding or removing a method is not backward compatible, modifying a body is, etc...).

This is how Hotswap works, the JVM has been using it for years.

As for changes that are backward incompatible, you can still manage them with application level techniques, such as rolling out servers or simply allow two different versions of the class to exist at the same time (JRebel does that, as do other a few other products in the JVM ecosystem).

Erlang doesn't really have any advantages over statically typed systems in the hot reload area, and its lack of static typing is a deal breaker for pretty much any serious production deployment.

> As for changes that are backward incompatible, you can still manage them with application level techniques, such as rolling out servers or simply allow two different versions of the class to exist at the same time (JRebel does that, as do other a few other products in the JVM ecosystem).

Neither of these allow for the whole reason Erlang has hot code upgrade in the first place: allowing to upgrade the code on one side of a TCP connection without dropping the connection to the other side. Tell me how to do that with a static type system :)

Tomcat (and other app servers) has support for doing hot reloads of Java web apps while not reloading the HTTP layer (and not dropping TCP connections).

http://www.javacodegeeks.com/2011/06/zero-downtime-deploymen...

I have implemented a similar system for JRuby apps running inside a Servlet container. There are many caveats. I don't actually recommend it because for a while you're using nearly twice the memory (and JRuby is particularly memory hungry). Also there are many ways to leak the old class definitions such that they are not GC'd (e.g. thread locals). But it's certainly possible.

I suspect that Erlang, Java, and all languages are in the same boat: some parts can be upgraded live in the VM while other parts require a full restart (maybe coordinating with multiple nodes and a load balancer to achieve zero-downtime).

Erlang is not in that boat. Generally, you can upgrade the entire thing (except the lowest level libraries obviously) without too much fuss.

Out of curiosity, where/why would such an exotic feature be needed in today's internet architectures where you always front a group of servers with a load balancer ?

Not all Internet protocols are HTTP. If you're running a service where long-lived connections are the norm, "simply fronting a bunch of servers with a load balancer" can require a pretty smart load balancer. E.g. IMAP connections often last hours or even days, and are required to maintain a degree of statefulness.

Not everything is a website! Also, not everything is stateless. Consider writing a chat application for the web for example and letting users on one page communicate with another one.

> lack of static typing is a deal breaker for pretty much any serious production deployment.

Are you talking about Google only where they made it a mandate or in general? There are serious production deployments on Python, Ruby, Erlang and Javascript.

I will trade expressiveness and less lines of code with a strong but dynamically typed language + tests over more a static typed language with more lines of code all being equal.

Or put it another way, if strong typing is the main thing that protects against lack of faults and crashes in production, there is a serious issue that needs to be addressed (just my 2 cents).

Go gives you Erlangs concurency model

There are a number of significant differences between Erlang's and Go's concurrency models: Asynchronous vs synchronous communication, per-thread vs per-process heaps, send to process vs send to channel.

Go has asynchronous communication and synchronous communication.

And the other things you mention are in practice not significant differences. The model both use is based on Hoares CSP and the same general ways of using apply. Some of the specifics must accomodate differences but those are differences of implementation not the general model.

No, they are different models [1]. Go does not have asynchronous communication. Bounded channels are still synchronous, because there is still synchronization happening; the consumer can't get too far behind the producer.

[l] https://en.wikipedia.org/wiki/Communicating_sequential_proce....

Go doesn't have strong type safety either; I remember a recent story about a Go stdlib function "accidentally" calling an interface it shouldn't.

it wasn't accidental -- it was written on purpose by a programmer (a conversion from Writer to WriteCloser). it was immediately acknowledged as an error and eventually may be caught by the standard code examining tool "vet".

What would the static analysis that "vet" is performing enforce to stop this? No interface-to-interface downcasts?

eventually, within the context of the `go vet` tool, the http://godoc.org/code.google.com/p/go.tools/go/types package may be used to analyze interface conversions ("I said I expected interface type A, but I'm using it as interface type B", which is unusual for go programs). i think that answers "yes" to your second question, but I'm not on the go team, so take my opinion with a grain of salt.

short of disallowing interface-to-interface casts, it is indicative of an error and should be vetted as such. the particular case I described earlier was covered by the Go 1.0 guarantee, so it had to be documented rather than fixed.

Hmm. Seems hard to do soundly in the presence of higher order control flow (e.g. pass the interface to a closure in a global variable—will the closure downcast it to another interface you didn't expect?)

Benefits in "performance" is somewhat vague. What performance criteria was measured? Throughput? Memory consumption? Latency? Long tails? Standard deviation? Mean response time?

It's highly unlikely that any language can possibly provide all the above and Erlang makes certain tradeoffs (as does Go). Low level languages like C++ let the user choose the tradeoff at any given point with enough effort.

What was Google's criteria in making such a decision?

Reply on news.ycombinator.com