asynclogger | 基于无锁队列（RingBuffer）实现的具备轮转功能的异步日志。

The random name 457e2f02 of your LoggerContext suggests that all your applications are using a single context. When one of your applications is stopped, the stop method is called on the context and no further messages can be logged. Meanwhile the reloaded application creates a context for itself.

This can occur, if you are using the Log4j API directly in your applications, but you don't ship the applications with a copy of log4j-api: in this situation the LogManager class in the shared Tomcat classloader will be used by all applications.

Try adding log4j-api.jar to your applications.

Since PostgreSQL driver does not support streaming results (as pointed out by @ave) you most likely want to implement result pagination. This is usually done either by using ORDER BY element_timestamp LIMIT OFFSET or by using keyset pagination. Either of these methods will load smaller batches of results into memory hopefully avoiding OOME by allowing GC to free each batch separately.

However it should be noted that sometimes increasing the available heap memory is the easiest solution. RAM is usually cheaper than programmers time so perhaps you should start by estimating how many records are you trying to process and how much memory is needed.

The message order bug was fixed in version 2.10.0. So from this version the messages should be displayed in the order. According to this answer there is only one thread that writes to file

That's because of this configuration where you have mentioned logger name as "logging".

You need to provide an ordering predicate in the keywords::order named parameter of the sink constructor. Given that you use logging::attribute_value_ordering, you should specify the attribute name based on which you want to order log records when constructing the predicate. For example:

You have given a wrong value in the additivity value of both loggers. Setting the additivity value to false will solve the messages duplication issue as it will prevent the logger from inheriting from its ancestor, in this case the root logger.

For the file name issue you mention in your comment: the filePattern property is used to generate the rolled log file name, that's the one on which you no longer write but keep for auditing/troubleshooting purposes, not the currently active one whose name is defined in the "fileName" tag instead.

I suspect that your application is producing log events faster than the Appenders can handle. The configuration shows two ConsoleAppenders. Logging to the console is extremely slow, literally 50x slower than logging to a file.

Async Loggers use a ringbuffer. If this buffer is full (because the Appender cannot take out events fast enough), then the application will not be able to add new events to the queue, so effectively the application is blocked from making progress.

My advice is to only log WARN or ERROR events to the Console, and everything else to files.

Another potential bottleneck is the custom Appender. Note that it opens a new FileWriter for every event. This is going to be slow. Better to open the file once and keep appending to it.

That error is just a warning output by the driver tests when byteman (a java agent) is not used. You can ignore it.

runPartyXServer is likely just starting the spring web app, not running the node itself.

Try running the nodes/runnodes.sh script or java -jar corda.jar in the directory of each generated node.