Phoenix performance issues with API

I’m using Phoenix to build a simple API that takes an HTTP request (JSON) and add the IP and user-agent to it and then insert it in MongoDB

 def imp(conn, params) do
    params=Map.put(params,"ip",conn.remote_ip |> Tuple.to_list |> Enum.join("."))
    params=Map.put(params,"ua",get_req_header(conn, "user-agent"))
    Mongo.insert_one(:mongo, "impression", params)
    json(conn,  %{})
  end

and it worked well, but what really is surprising is the performance, I have implemented the same function using NodeJS and this was the responses time of both:

                 Node     Phoenix 
#1             237 ms       4s
#2             65   ms      571ms
#3             30   ms     575ms
#4             33   ms     581ms
#5             27   ms     577ms
#6             30   ms     571ms
#7             31   ms     567ms

I have read about phoenix performance and how it should respond in microseconds time.

I thought that dealing with MongoDB is the issue, so I delete the insertion code but I still get 953ms response time (the first time) and 89ms, 350ms, 69ms the other times

so why my phoenix app is not that fast?

Note: I’m using Windows and Postman

are you running phoenix in dev or in prod?

1 Like

I’m running phoenix in dev, but really is it that much slow in dev?, I’m running Node in dev as well and it is not that slow

yep, much slower - also why the first request is so slow (it compiles on first request)…

for any benchmarking run in prod…

1 Like

Benchmarks should always be executed against a production mode. Development mode is very different, and it is setup for development convenience: when writing code you want changes to appear in your app without having to restart it, and other convenient things, and you normally don’t care about microsecond response time.

3 Likes

code_reloader: true in dev.exs can cause this, try setting it to false. But you shouldn’t test response times in dev mode any way. Also I’m sure that almost no one uses Windows in production so it might not be up to par with with Unix/Linux. If you want to test response times in Windows I would use WSL (Windows Subsystem for Linux) and run your app there in prod mode.

7 Likes

Thank all of you guys.
Actually, your responses are relieving

I have tried to set code_reloader: false and I started to get 30 ms responses and I will benchmark the app in production mode next times

6 Likes

code variations for benchmarking:
pipe the map.puts (there might more effective ways than map.put)
avoid json encoding since you are not returning anything
try other “ip to string” method (which btw also works for ipv6 down the line)

 def imp(conn, params) do
    insert=
    params
    |> Map.put("ip",conn.remote_ip |> Tuple.to_list |> Enum.join("."))
    |> Map.put("ua",get_req_header(conn, "user-agent"))

    Mongo.insert_one(:mongo, "impression", insert)

    conn
    |> put_resp_content_type("application/json", nil) #necessary?
    |> send_resp(200, "{}")
  end

also try conn.remote_ip |> :inet.ntoa() |> to_string() and see if it’s faster, also your current code will break on ipv6 afaik…

since this seems to be a stats endpoint consider wrapping the db call in task.start, which will of course void the benchmark as endpoint will now return crazy fast - but this is what I do and would somewhat recommend for this kind of endpoint…

Task.start(fn -> 
  Mongo.insert_one(:mongo, "impression", 
    params
    |> Map.put("ip",conn.remote_ip |> Tuple.to_list |> Enum.join("."))
    |> Map.put("ua",get_req_header(conn, "user-agent"))
   )
end)
1 Like

Thanks for this amazing reply, but could I ask about what are the other effective ways should I use because I’m using Map.put all over my code

personally I didn’t know any, but a bit of googling shows me to use map.merge instead of multiple “map.put” see/read https://joyofelixir.com/10-maps/

1 Like

Consider that we’re getting in the realm of micro-optimizations here. At the point when you add a call to MongoDB, especially if synchronous, it probably does not make sense to optimize Map.put vs. Map.merge and rather use whatever is more readable, as the database IO would dominate the time.

4 Likes

Back when I still used Windows 10 for work (stopped September 2019) Erlang and thus Elixir were definitely slower than on Linux and macOS. Just one extra data point.

Even on a pretty pathetic laptop (i3 CPU with 2 cores) my dev mode apps with code reloading still respond in 5 to 50 ms in Linux. And on my main workstation (macOS) I often see response times below 1ms.

2 Likes

Definitely test in prod like others already suggested. Also make sure that not too much logging occurs, e.g. by setting the log level to warn. There are a few other minor tips which can improve numbers. I blogged about it here. The article is 4 years old, but hopefully most of the tips still apply :slight_smile:

9 Likes