Design – Is it worth logging HTTP requests when they enter an API server

You're right in that this question is very much answered by, "what are your requirements", but this is my opinion.

Be sure that any event initiated by a user is traceable back to the user. Each logged event should state who is executing the event, or at least contains a transaction ID that can be traced back to the user creating the transaction. This will allow you to fulfill non-repudiation requirements where a user would not be able to deny creating the transaction.

I wouldn't recommend creating a new log file for each session. A better approach would be to create new log files once they reach a certain size (10 MB for instance). Or, just creating a new log file each day or week (depending on the size of the files). If there a lot of events being logged, a database may be better as it can handle large volumes of data easier and you will be able to query the results as needed.

For libraries, there are plenty out there for Java EE. Log 4J is fine. I've also used Apache's Commons Logging, but I prefer the combination of SLF4J and Logback personally - they are easy to use and integrate well with other logging frameworks.

You may also want to look into Aspect Oriented Programming to create the common logging events that will appear in a lot of methods. Specifically, AOP can help create debug logging for method enter / exit / throw. I've done AOP with Spring and can say it is very helpful keeping the code clean.

Speaking of logging levels, it will be very beneficial to make good use of each level as appropriate and create a secure way to change the level at runtime if you are experiencing issues.

Log only errors that would be beneficial to have logged. Tautological, maybe, but logging can often get out of hand. Having too many records to sift through means you will not be able to identify outliers or actual problems. Logging too little means you may miss an issue (if you didn't log it, it never happened), so, make good use of logging levels for these reasons as well. There may be no reason to log a user validation error (e.g., creating a new password less than the allowable number of chacters), but system errors should always be logged (e.g., unable to contact the database).

Likewise, only notify the user of what they need to know. They will need to know if the password they entered failed validation and why, but they don't need to know if the database is down (they only need to know that the site is experiencing issues and what they can do in the mean time - i.e., call a support number). Chattiness to the user should be kept to a minimum as constant error notifications that they can't personally resolve would be a poor user experience.

I suggest you take a look at OWASP's Logging Cheat Sheet for some more information on logging best practices.

Return the harshest, unfriendliest result possible in the event of a malformed request (one that returns more data than your metering allows is malformed). I suggest returning a 4** error code. Then, also provide paging parameters, so that users may request pages. oData has this feature, for instance. Do not truncate the data silently, under any circumstances.

Consulting with customers is a bad idea. They are going to tell you to do whatever possible to minimize errors, which is a bad engineering approach. This is your decision, take it by the horns and do the right thing.

An example of a paginated api is oData:

http://www.odata.org/documentation/odata-version-2-0/uri-conventions/