[GNC-dev] Bug 797463 - CSV Import of transactions into a new file hangs

[Christian Gruber]

Am 11.11.19 um 11:22 schrieb Geert Janssens:
> Op zondag 10 november 2019 23:28:06 CET schreef Christian Gruber:
>> Am 10.11.19 um 22:39 schrieb John Ralls:
>>>> On Nov 10, 2019, at 12:07 PM, Christian Gruber
>>>> <christian.gruber at posteo.de> wrote:>>
>>>> Am 08.11.19 um 23:08 schrieb John Ralls:
>>>>>> On Nov 8, 2019, at 1:58 PM, Christian Gruber
>>>>>> <christian.gruber at posteo.de> wrote:>>>>
>>>>>> Am 08.11.19 um 04:39 schrieb John Ralls:
>>>>>>> Christian,
>>>>>>>
>>>>>>> It's not that it's not prepared for Bayesian matching, it's that older
>>>>>>> versions of GnuCash stored the Bayesian match tokens hierarchically.
>>>>>>> Aaron Laws (lmat) changed it to a flatter structure with somewhat
>>>>>>> better memory locality for faster access. imap_convert_bayes_to_flat
>>>>>>> should run once to convert the data and set the feature, after which
>>>>>>> check_import_map_data will see the flag and return. A file created
>>>>>>> with 3.x and Baysian maps would already have the feature set.
>>>>>>>
>>>>>>>
>>>>>>> With that background, to your questions:
>>>>>>>
>>>>>>> Why does it take so long? Because it traverses the entire tree of
>>>>>>> accounts, every time. The test book has 1127 accounts. Add to that
>>>>>>> that there are some things inside the loop that shouldn't be and that
>>>>>>> convert_imap_account_bayes_to_flat doesn't use some obvious short
>>>>>>> circuits and you get taking a long time.
>>>>>>>
>>>>>>> Why does it run twice? Because there aren't any accounts with
>>>>>>> import-map-bayes slots, so it does no conversions so it doesn't set
>>>>>>> the feature.>>>>
>>>>>> Why isn't the feature set in any case after conversion is done, whether
>>>>>> there was any slot to convert or not?
>>>>>>
>>>>>> I would expect that, if the hierarchical structure should not be used
>>>>>> anymore. I wouldn't only set the feature, if there has been any error
>>>>>> during conversion. But it's not an error, if
>>>>>> convert_imap_account_bayes_to_flat() returns false.>>>
>>>>> The feature isn't set because that would prevent using the file with an
>>>>> older version of GnuCash for no good reason: There aren't any
>>>>> import-map-bayes tags in the new format so there's nothing for the
>>>>> older version to mess up.
>>>>>
>>>>> Regards,
>>>>> John Ralls
>>>> Sounds meaningful, but it's not obvious. The conversion process is not
>>>> much transparent to the user. The user is not informed about conversion
>>>> and additionally the conditions, when the GnuCash file is converted and
>>>> when not, are not quite intuitive, I guess. And the second question for
>>>> me is, if this is really a relevant use case to keep the GnuCash file
>>>> unconverted if possible, when it is used with a newer GnuCash version.
>>>> Maybe it's relevant, if several people, who use different GnuCash
>>>> versions work on the same GnuCash file. But what happens, when one with
>>>> a newer GnuCash version uses the bayesian import matcher for the first
>>>> time, which silently converts the GnuCash file? Will it still be usable
>>>> with an older GnuCash version?
>>>>
>>>> But actually this is another discussion, which is not important for the
>>>> bug ticket. What's more important, I created a new fresh GnuCash file
>>>> with the SKR04 accounts template using my GnuCash version 3.7 and if I
>>>> checked this correctly, even with this GnuCash version the feature
>>>> GNC_FEATURE_GUID_FLAT_BAYESIAN is not set in the GnuCash file. Can you
>>>> confirm this? If yes, then the bug report is relevant not only for
>>>> migration from older to current GnuCash versions but even for current
>>>> GnuCash versions in general.>
>>> No, it's the same discussion. If you don't use the feature then the flag
>>> for it won't be set. That's how we designed the feature mechanism. The
>>> use case is straightforward: Some user has two computers, a laptop
>>> running Windows and a desktop running Ubuntu 16.04. She updates the
>>> Windows machine with each release but doesn't know how to build GnuCash
>>> on her Ubuntu machine so leaves it at whatever they shipped on 16.04;
>>> 2.6.12 IIRC. As long as she doesn't use any of the features supported
>>> only in 3.x she can continue to use both systems to keep her books and
>>> regardless of which machine she uses to create a book.
>>>
>>> The "silently" part could be better. It would indeed be a nice enhancement
>>> to pop a dialog box when GnuCash is about to set a feature flag
>>> explaining that it's use will preclude using the file on a GnuCash
>>> version older than X and giving the user a chance to bail out. It would
>>> be nicer still to persist that decision so that it need ask only once,
>>> and if the user assents it could set the feature flag even if the feature
>>> isn't used, as is the case here. Of course there would then be the need
>>> to reset negative responses so that when our hypothetical user upgrades
>>> her desktop to Ubuntu 20.04 next spring she can turn on all of those 3.x
>>> only features she'd previously declined.
>>>
>>> It occurs to me that another improvement to the existing flat Bayesian
>>> conversion would be to check the state of the GNC_PREF_USE_BAYES and skip
>>> the conversion if it's not true.
>> This is already implemented. Flag GNC_PREF_USE_BAYES is checked in
>> functions matchmap_find_destination() and matchmap_store_destination()
>> and if not set, than functions gnc_account_imap_find_account_bayes() and
>> gnc_account_imap_add_account_bayes() aren't invoked and no conversion
>> will be done.
>>
>>> Regards,
>>> John Ralls
>> Thanks for the revealing explanation of the feature mechanism. In this
>> case it would be really helpful, if GnuCash would explain itself and
>> interact with the user, when features are changed and conversions are
>> done, which are not backwards compatible. Moreover it would be
>> meaningful, if the user could "enable/disable" backwards compatibility
>> when creating a new GnuCash file to avoid such extensive conversion
>> steps. This means, that all supported features of the used GnuCash
>> version are set by default during creation. Alternatively, if it's
>> possible to significantly speed up searching for (non-existing) bayesian
>> match tokens, this should be preferred.
>>
>> I'll create an enhancement ticket.
>>
>> Regards,
>> Christian
> I fully agree there should at least be visual feedback to the user when
> running the conversions. There are technical challenges in accomplishing that
> in this case though.
>
> Another potential performance improvement would be to ensure the conversion
> will run only once per session, even if the feature flag doesn't get set. At
> least that would mean the slowdown happens only the first time one tries to
> import during a longer session.
>
> Regards,
>
> Geert
>
>
I thought again about the case, that the user confirms conversion, but 
the conversion is actually not necessary, since there was nothing to 
convert like in this case.

This could also be seen as an implementation problem. If GnuCash could 
exactly determine, when a conversion will be necessary, before asking 
the user, then the problem would be solved.

But in this case it's actually not only the conversion itself, which 
needs extensive processing time. It is even the process of checking, if 
a conversion would be necessary or not. So one solution could also be, 
that GnuCash somehow stores the result of this check, which is different 
from setting the feature flag GNC_FEATURE_GUID_FLAT_BAYESIAN.

Futhermore, if GnuCash should check, if the conversion is really 
necessary, before asking the user, the current implementation has to be 
changed before adding user interaction, since both the check and the 
conversion itself are implemented in the same function 
imap_convert_bayes_to_flat() currently.

On the other hand it's not that bad to ask the user even in uncertain 
cases, i.e. when it's only assumed, that a conversion will be necessary. 
But if the user then confirms conversion, it must be performed, i.e. the 
feature flag has to be set anyway.

Regards,
Christian