Recently at a client site, we encountered an error that was a little more baffling than normal.  Under very specific circumstances,    RLIST would bail with no results.  Eventually, we had a duplicatable test sentence which was basically

    SELECT TABLE BY CALC_COL_1 WITH REAL_FIELD_1 = ‘XXX’ AND WITH CALC_COL_2 = ‘YYY’

When we removed the BY clause (CALC_COL_1) our sentence ran to completion.   This put the problem squarely on CALC_COL_1, which is where we started debugging.  

We changed CALC_COL_1 to a simple “@ANS=1” and ran the sentence again.  This worked correctly.  To prove to ourselves that CALC_COL_1 was the problem, we put back the original code and changed our sentence to 

    SELECT TABLE BY CALC_COL_1 WITH REAL_FIELD_1 = ‘XXX’ AND WITH CALC_COL_3= ‘ZZZ’

This also worked.   

Now we’re starting to get a little confused.  Modifying CALC_COL_1 solved the problem.  Replacing CALC_COL_2 also solved the problem.  It has to be some sort of interaction with the calculated columns, because there’s no obvious consistency lining up the errors with the sentence structure.

It’s time to break out the profile log and see if something jumps out, and it did.  Something appeared off with EXTERNAL.SORT.  It seems like it executed a few times, and then everything stopped.  We dusted off our copy of PROCMON.EXE ¹ and ran the process again.  PROCMON showed that the sort file was being opened, but that it was not being closed.  It also showed a lot of file access  We’re now convinced this has something to do with sorting in general, and not with the specific field being sorted.

After trying well over a dozen different variations on the RLIST sentences, we finally worked out that we needed two things to happen for the error to occur.  The first was that we had to have CALC_COL_1 as the sort column.  The second was that the WITH clauses needed to return enough keys to force the system into using EXTERNAL.SORT instead of a simple in-memory sort.

Normally, when OpenInsight starts behaving very oddly, the culprit is usually a function that’s not clearing set_status() before calling get_status.  When that happens, the get_status() call could be returned an error from 10 programs back.  We reworked the calculated columns to ensure we clear set_status, and feeling very proud of ourselves, sat back and waited to bask in the glory offered to us by our client.  Obviously, the report failed again.

Sitting back, we start wondering if maybe there’s a small part of the RLIST code that’s still thinks it’s Advanced Revelation and it’s actually checking the pseudo STATUS() variable.  So, we reworked the calculated columns to ensure we clear set_status and status(), and feeling very proud of ourselves, sat back and waited to bask in the glory offered to us by our client.  Obviously, the report failed yet again.
 
Surely there can’t be three ways to set an error condition in OpenInsight?  It turns out that there is.  We suddenly remembered @FILE.ERROR.  So, we rework the calculated columns to ensure we clear set_status and status() and @FILE.ERROR, and feeling very proud of ourselves, sat back and waited to bask in the glory offered to us by our client.  And this time there was much rejoicing and basking.

With the solution in hand, we set out to find out exactly why this happened and where the system was registering the failure.  What we found was the calculated field was doing a lot of file access and would attempt to access records that didn’t exist.  This was what was setting @FILE.ERROR.  But, we hit a two-fer, as they say.  Some of the randomness had to do with the amount of data returned.  If enough sort data was returned the system would call EXTERNAL.SORT to sort the data.  EXTERNAL.SORT writes to temporary OS files, and the OS file commands use @FILE.ERROR to check for errors.  EXTERNAL.SORT would execute an OS operation, check @FILE.ERROR, find the pre-existing error, and abort the process.  But, this would only happen if the very last row processed set @FILE.ERROR.  Otherwise RTP57 would clear any prior @FILE.ERROR settings, and @FILE.ERROR would be null during the EXTERNAL.SORT call. 

1.  PROCMON and other useful tools are available as part of the Sysinternals suite of development tools, which no programmer should be without.

Those of you who've been in the industry for a while will remember the smugness with which we treated the Y2K issues the rest of the world were experiencing. It didn't affect us, we stored our dates in internal numerical format. I'm sure we all have favourite anecdotes. Mine was being called into the Press Office of a major Government department to certificate their AREV tracking system. I sat down and gave the system a cursory glance. "Ermmm - it doesn't seem to use dates?". "No, it doesn't". "Well, here's your certificate"... turns out they couldn't get the budget for system tweaks but they could for Y2K compliance. I spent the rest of the day implementing their desired changes.

Anyway, this week wiped the smile off of our collective faces (if you'd ignored the advice in KB 42 thirty or so years ago - which to be honest you could be forgiven for) with the advent of our very own variant - we'll call it the 20100 bug, although it's not a bug, it's an unfortunate feature.

Users began reporting that date searches were failing for values after the 10th of January 2023. At first we couldn't see an obvious reason. We built a database containing date values going back to the previous century with five rows per date and indexed it. We wrote a test program and ran it

RUN ZZ_POP 10/01/2023

RUN ZZ_POP 20099

What we've fallen foul of is discussed in the KB reference earlier. Basically BTREE.EXTRACT takes what it is given and tries to ICONV it to use for the look up. If it can't ICONV it (20099 isn't a viable date) it uses the value passed. If it CAN ICONV it, is uses the ICONVed value. And guess what? 20100 ICONVs to the 2nd of January 2000 as you can see above.

Of course, this won't be the first time this has happened in the wild - for example looking for the 7th of December 1998 using 11299 would have returned the 1st of December 1999 and so on.

Note that the same issue will be experienced when using internal date formats with RLIST statements.

The solution? When calling BTREE.EXTRACT use EXTERNAL values as it will try and ICONV the data before using it.

Addendum

We've been asked to share the code we used to locate btree.extracts in client systems to enable a manual inspection to determine possible failure points. This is a rough and dirty hack which met our requirements. Feel free to tailor to your own requirements.

Tabbed interfaces provide a much loved way of fitting lots of information onto a single entry form. The introduction of multi page forms and the improved integration of page controlling made this a much easier exercise - each tab could be a page of the form and we could simply move between pages on a tab click. In fact a quick event was added to make this even easier.

But what has not been possible until now, without a deal of smoke and mirrors, is to have one or more tab controls embedded within a page on a single or multi-page form and have them only appear on a specific page. The only way to accomplish this would be to have a set of controls per tab, all within the tab boundaries, and make them visible or invisible depending on the currently clicked tab. This makes form maintenance incredibly tedious and code overly complex.

Before OI10 there was no real concept of parent/child relationships in OI - other than that forms and group boxes owned all of the controls within their boundaries. This led to many issues, not least with group boxes and tab controls. With the advent of OI10 this has changed. OI now allows the implementation of, and respects, parent/child relationships.

The biggest example of this is the introduction of a "Panel" container type.

A Panel is simply a container that can be placed anywhere on the screen so that when controls are pasted into the Panel they become children of the Panel, not of the Window. So move the Panel and the controls owned by the Panel move with it. Don't be too put out by the "Simple" part of the description, pretty much the only difference between a Panel and a Simple Panel has to do with paging. So looking at Simple Panel/Panel  properties in the Form Designer we see this :

    PageCount
    Virtual Size
    PageSwapRenderMode
    ScrollMode
    TabWithinPanel

and the only property unique to Simple Panels is the DummyCaption property.

So, to implement a tab within a form page, we can now combine the Tab Control, the Panel Control and one simple Quick Event to achieve the desired results.

Let's start with a simple one page entry screen.

(As an aside, if you right-click on the Window you can un-anchor it to allow moving it wherever you want.)

Tell the Panel that it has three pages and label the tab accordingly (after setting the tab to three values) :

Select the Panel and add in the controls for the first page.

Addendum - Mouse Use In Form Designer