Difference between revisions of "User:Lmat"

Revision as of 09:04, 5 December 2017

Bayes GUID Work

The Old Plan

In the next 2.7 release (2.7.3?), if GNC_FEATURE_GUID_BAYESIAN is not set, a conversion will take place to change all the import maps to guid and flat^[1]. Only this 2.7 release will set the feature flat.

The next 2.6 release (2.6.19?), when GNC_FEATURE_GUID_BAYESIAN is set, all reads and writes will be done in the new style. If it is not, all reads and writes will be in the old style.

The Problem

If books are created in 2.7.2, then bayes entries are added in 2.6.18, the import map will be mixed. If this is then read by either the new 2.7 release or the new 2.6 release, they will fail to get all information.

The New Plan

The next 2.7 release (2.7.3?) should have a new feature: GNC_FEATURE_FLAT_GUID_BAYESIAN. It will only be set by this and subsequent releases. When it is set, the import maps will be converted.

The next 2.6 release (2.6.19?) will be able to read the feature GNC_FEATURE_FLAT_GUID_BAYESIAN.

If books are created in 2.7.2, then edited in 2.6.18 to make a mixed-import-map situation, then loaded in the new 2.7 release, what should happen?

gjanssens: When the feature flag protection is properly implemented this can't happen. If the books are created with 2.7.3 and the user has created bayesian imap entries in that version, GNC_FEATURE_FLAT_GUID_BAYESIAN will be set. When you then try to open that book with 2.6.18, it will refuse to do so because 2.6.18 doesn't know this feature. So in this case 2.6.18 can't create a mixed-import-map situation any more. If on the other hand 2.7.2 didn't add guid based import maps the feature flag won't be set by 2.7.2. This means gnucash 2.6.18 will be able to open this file as there is no data format conflict. If subsequently the user adds import maps in 2.6.18, they will all be in the old format. A subsequent run of 2.7.2 on that file will then find account name based import maps in in it while GNC_FEATURE_FLAT_GUID_BAYESIAN not set. Hence it will run a conversion of all import maps and set the feature flag. From that point on the file will no longer be opened by 2.6.18 (full circle). Note it's crucial here the feature flag should only be set if there actually are flat/guid based import maps in the data file. Otherwise you'd be preventing 2.6.18 from reading the file even if it didn't have any import maps at all. Yet in that case the file is perfectly readable by 2.6.18 and the user should be allowed to do so.

jralls: Not quite. If the file is created in 2.7.0 - 2.7.2 and the user did an import with bayes matching then GNC_FEATURE_GUID_BAYESIAN (no FLAT) will be set; if the file was created in an earlier version and has bayes import maps with account names then those get converted and GNC_FEATURE_GUID_BAYESIAN is set and the account name entries are converted to guids. Trying to open either file with Gnucash older than 2.6.12 will fail because the feature is unknown. Opening the file with 2.6.12-2.6.18 will read the guids OK and if it needs to write new ones will clear the feature flag so that opening it with 2.7.0-2.7.2 will know to convert the new entries and reset the feature flag.

So 2.7.3 and 2.6.19 will need to handle three possible situations: 1)GNC_FEATURE_GUID_BAYESIAN is not set, where some or all of the entries have account names instead of guids. Those that do need to have the account names changed to guids and all entries need to be flattened. 2) GNC_FEATURE_GUID_BAYESIAN is set, so all of the entries will have guids and will need only to be flattened. 3) GNC_FEATURE_FLAT_GUID_BAYESIAN is set, no conversion required.

2.6.19 should write account names and not-flat if neither flag is set and guids and flat if GNC_FEATURE_FLAT_GUID_BAYESIAN is set. Where GNC_FEATURE_GUID_BAYESIAN is set it could either unset it and write account names and not flat or leave it and write guid and not flat.

Once the GNC_FEATURE_FLAT_GUID_BAYESIAN flag is set by 2.7.3 neither 2.7.0-2.7.2 nor anything older than 2.6.19 will be able to open it for writing.

Thank you for your replies. The implication thus far is that the answer to my present question is "yes", but I would like to make it explicit: is it important that we support data files from unstable versions? Lmat (talk) 01:17, 5 December 2017 (UTC)

Well, your question was "what should happen?" and I think I laid that out for both versions. Whether it's worthwhile to implement all three cases is a bit of a probability game. It's entirely possible that there are zero instances of case 2 or of case 1b, where GNC_FEATURE_GUID_BAYESIAN isn't set but some of the entries are still already in guid form. You can finesse the latter by including instructions that if there's Bayesian data and it's ever been opened by 2.7.[012] that one needs to open it in 2.7.[012] and save before opening it in 2.7.3. Or we can say that since we warned about not using production data that any book that's been opened with 2.7.[012] should be discarded if it has Bayes matching data. Or handle all 3 (4?) cases. All are acceptable options, but we have to pick one and if it't not the last one we have to say what to do in the release notes. John Ralls (talk) 03:46, 5 December 2017 (UTC)

My reasoning silently assumed we don't support data files manipulated by beta versions. If we can with a reasonable effort, that would be the most user-friendly way. It seems to me supporting all 3/4 uses cases adds quite some complexity in both 2.6.19 and the 2.8 series which we have to carry around for quite some time. I haven't looked too deeply in the code though. Second best would be to add some code that detects unsupported situations and tell the user how to resolve it if it's too complicated to write code for all possible conversions. The simplest form is to consider files with import maps manipulated with 2.7.[0-2] as unsupported as this feature only existed in early betas of the 2.8 development cycle. In that case we could even drop GNC_FEATURE_GUID_BAYESIAN from 2.7.3 as well, so anything more recent will refuse to open these files as well. That should be accompanied with a clear release note message though. I'm having a hard time deciding which is best. Perhaps we should poll our user base to get an idea of how many people discarded our advice ? Gjanssens (talk) 09:04, 5 December 2017 (UTC)