Httpclient大量的close_wait

As I discovered when debugging TCP connections stuck in the CLOSE_WAIT state for one of our customers, we were using HttpClient incorrectly. We’re not alone in this case, as you’ll find out if you google HttpClient CLOSE_WAIT, but it’s very non-intuitive. Even the official tutorial is wrong, so I’m describing the issue here.

Apache HttpClient is usually used like this in basic mode:

HttpClient httpClient = new HttpClient();HttpMethod method = new GetMethod(uri);try {int statusCode = httpClient.executeMethod(method);byte[] responseBody = method.getResponseBody();// ...return stuff;} finally {method.releaseConnection();}

But this is not enough.

The issue is that releasing the connection makes it available again to the HttpClient instance, but does not close it, because HTTP 1.1 is used and it can pipeline further requests to the same host:port in the same connection.

Even though the server may have decided to close its end of the connection, on our client side the connection is still open and will stay that way until an attempt to read from it is made (at which point the client will detect that the other end is closed). TCP works like that, there is a notion of half-closed connection, because close() actually just means I will not send any more data, and you can still receive data from a connection that you closed but which has not yet been closed on the other end.

So what happens then is that when the HttpClient instance goes out of scope it becomes available to the GC, but it will not be garbage collected immediately. Until the GC collects it, the socket connection held internally will stay open and the socket will be stuck in the CLOSE_WAIT state.

To fix this, the simplest way is to add:

method.setRequestHeader("Connection", "close");

before executing the method. This will instruct HttpClient to close the connection by itself once the full response has been received. Another way is to do in the finally block:

httpClient.getHttpConnectionManager().closeIdleConnections(0);

An even better way is to not use a new HttpClient object each time, but to reuse one that has been initialized with a MultiThreadedHttpConnectionManager sized appropriately. Of course in this case the connection manager must be shut down properly when the application shuts down:

private MultiThreadedHttpConnectionManager connectionManager;private HttpClient httpClient;public void init() {connectionManager = new MultiThreadedHttpConnectionManager()// ... configure connectionManager ...httpClient = new HttpClient(connectionManager);}public void shutdown() {connectionManager.shutdown();}public String process(String uri) {HttpMethod method = new GetMethod(uri);try {int statusCode = httpClient.executeMethod(method);byte[] responseBody = method.getResponseBody();// ...return stuff;} finally {method.releaseConnection();}}

Florent

PS: I’m using here the APIs from HttpClient 3, but it also applies with slightly different names to the completely refactored APIs of HttpClient 4.