Dale, Thanks for doing this work and especially for adding these guidelines to the Wiki!
On Thu, Nov 4, 2021 at 1:08 PM Dale Emery <dem...@vmware.com> wrote: > As of July, 2021, Geode's build system no longer executes test worker > processes in separate Docker containers. > > This increases the risk of port collisions between tests. Each test worker > JVM and each Java process started by a test executes directly in the > environment provided by the host machine. If a test is using a port, any > concurrently running test that tries to bind to the same port number will > suffer a BindException. > Safely allocating ports. To reduce this risk, Geode's build system now > assigns each test worker JVM a distinct range of ports. In CI, which runs > 24 distributed tests concurrently, each test gets a range of about 400 > ports. > In JVMs running Geode version 1.14 or later, AvailablePort and > AvailablePortHelper will choose ports from this assigned range. As a > result, no two tests will ever receive the same port number from these > classes, no matter how many are running concurrently. > When you write a test, keep the following guidelines in mind, to make sure > your test uses only the ranges of ports that the Geode build system > assigned to it: > > * To assign an available port, call AvailablePortHelper or > AvailablePort. These are the only classes that know about the reduced port > range for the test. > * Call AvailablePort and AvailablePortHelper only in the current > version of Geode. Do not call these classes in Child VMs running older > versions of Geode. Older versions of Geode do not honor the new, reduced > port ranges. > * Do not launch any service using Geode’s default port for that > service. Always explicitly assign an available port. The only safe use of > default ports is in a test to verify that a service uses its default port > by default. Note that such a test will fail the stress test CI job. To > exclude such a test from the stress test job, annotate it with > @Category(IgnoreInRepeatTestTasks.class). > * Do not launch any service using a port number hard-coded into the > test. Always explicitly assign an available port. There is no safe use of > any hard-coded port number in tests that can run concurrently. > * Do not attempt to reuse an ephemeral port. Some tests start a member > on an ephemeral port, stop the member, and attempt to restart it on the > same port. This is not safe. It was never safe. During the time between > when the member stops and when it restarts, the port is available for the > operating system to give to any other process. In CI, it is very, very > likely that one or more concurrently-running tests (or other processes) > will request ephemeral ports during the time when your test is not bound to > it. If one of those processes gets the port your test was using, your test > will fail. > This information is available on the Geode wiki: > https://cwiki.apache.org/confluence/display/GEODE/Using+Ports+Safely+in+Tests > > Dale > >
