Describe the bug
This is a follow-up to the issue #1233, and the last comment in particular.
I still encounter some issues with an opened Socket, but it's random and I didn't manage to come up with a simple reproducer.
To Reproduce
I did some investigations, and here are the findings:
DatabricksClientConfiguratorManager#instances is always empty after your fix, so that's good.DatabricksHttpClientFactory#instances is not always empty; sometimes there is still a DatabricksHttpClient with type TELEMETRY.
I suspect there is a race condition where the connection closes and calls DatabricksHttpClientFactory#removeClient(context), and afterwards a different thread calls DatabricksHttpClientFactory#getClient(context, TELEMETRY) with that connection's context.
I had to do this hack:
try {
final Field instancesField = DatabricksHttpClientFactory.class.getDeclaredField("instances");
instancesField.setAccessible(true);
final Map<SimpleEntry<String, HttpClientType>, DatabricksHttpClient> openedClients =
(Map<SimpleEntry<String, HttpClientType>, DatabricksHttpClient>)
instancesField.get(DatabricksHttpClientFactory.getInstance());
openedClients
.values()
.forEach(
databricksHttpClient -> {
try {
databricksHttpClient.close();
} catch (final IOException e) {
// Do not throw
}
});
} catch (final Exception e) {
// Do nothing and hope checkpoint can happen
}
Expected behavior
I would expect all the sockets to be closed when all connections have been closed.
Client Environment (please complete the following information):
- OS: [e.g. Windows] Ubuntu
- Java version [e.g. Java 21] Java 21
- Java vendor [e.g. OpenJDK] Azul
- Driver Version [e.g. 3.1.1] 3.3.1
Additional context
Add any other context about the problem here.
Describe the bug
This is a follow-up to the issue #1233, and the last comment in particular.
I still encounter some issues with an opened Socket, but it's random and I didn't manage to come up with a simple reproducer.
To Reproduce
I did some investigations, and here are the findings:
DatabricksClientConfiguratorManager#instancesis always empty after your fix, so that's good.DatabricksHttpClientFactory#instancesis not always empty; sometimes there is still a DatabricksHttpClient with type TELEMETRY.I suspect there is a race condition where the connection closes and calls
DatabricksHttpClientFactory#removeClient(context), and afterwards a different thread callsDatabricksHttpClientFactory#getClient(context, TELEMETRY)with that connection's context.I had to do this hack:
Expected behavior
I would expect all the sockets to be closed when all connections have been closed.
Client Environment (please complete the following information):
Additional context
Add any other context about the problem here.