There's a parameter on the CRTPF command known as the "force-write ratio." It lets you tell the system how many adds or updates are processed before they're written out to disk. I've never changed it when creating a file, and at the shops where I've worked, the default has always been set to *NONE. In typical IBM-ese, *NONE doesn't mean "none"; it means more than one, a whole lot, or as many as the system thinks you deserve. The point is that if your force-write ratio is anything other than one, you may not find your recent database changes when you look for them. That can be annoying. If the system hasn't decided to write them to disk, they must still be in "the buffer." Buffering database changes (and reads) is also referred to as "record blocking."

Of course, there is a reason why database changes are held in the buffer, and that reason is performance. Keeping disk I/O to a minimum makes your application run faster. And most of the time, buffering doesn't create a problem. Suppose your job has CL program A calling RPG program B to create a work file and RPG program C to process the work file and create a report.
When program B ends and program C begins, every work file record created by program B has been written to disk. You'll get the same results on your report whether the system wrote those records to disk in blocks of 10, 100, or 1,000. We trust the system to choose a blocking size that will optimize performance.

It's important to note that our example program works because program B ends. Since the program is no longer in memory, the work file it was using was either explicitly closed or implicitly closed when *INLR was set on. The buffer is always flushed to disk when a file is closed...which is why we're all not pulling our hair out every night trying to find those "missing records" in our daily batch cycles.

But in many situations, it's necessary to keep a program in memory. It could be a constantly running job that monitors a data queue or a server program that's being called repeatedly. These types of programs frequently remain resident in memory to avoid the overhead associated with being repeatedly loaded and unloaded. And a program that does a return when called from the command line for testing or debugging will also stay resident in memory (though perhaps not intentionally).

In these cases, it's possible that your program could write to a file, and those changes could be "stuck" in the buffer. In addition to the program remaining resident, a few other things must be true as well:

• The file you're writing to is output only.
• The file is not closed after processing and prior to the return statement.
• You haven't explicitly told your program that it shouldn't block records.

If you've determined that your missing records might be hiding in the buffer and your program needs to stay in memory, there are several ways to force the records to disk.

1. Use the Override Database (OVRDBF) command to override the force-write ratio of the file to 1. This will effectively turn off record blocking.

2. Define what you consider to be a logical transaction for your program. Open the file at the start of each transaction and close the file at the end of the transaction. This allows your program to take advantage of some record blocking but still ensures that records in your output file are available to any downstream process.

4. Trick the system into thinking that you're using the file for both input and output. Any random access will result in every database change going straight to disk. I don't recommend this, but I've seen it done.

Record blocking is great for performance, and most of the time, we can let the system handle it for us. But occasionally we need to be able to control what's in the buffer. So if you've got some missing records and you're at the point that you'd swear it's the compiler, some missing PTFs, or even sunspots, then maybe they're in the buffer! Hopefully, this will spare at least a few of you some long hours of staring at your terminals.

Note: Although IBM says that this can happen on updates as well as adds, I've never known that to be the case. In testing, it was easy to create a situation in which records remained buffered when adding records. But I was unable to create the situation for updates. If anyone has ever seen that, I'd like to know about it.

Mike Savino currently supports J.D. Edwards WORLD Financials and Order Entry for the Consumer Healthcare Division of Wyeth.