My apologies for the elements of review in what follows. Writing them helped me organize my thoughts, so I hope that reading them will help communicate those
As Herbert reminds us, for just about any category that might appear in a data file, one can imagine an experiment, a construct, a model, etc. whose description
requires multiple instances of that category. As James observes, however, many categories in our current dictionaries so rarely require such treatment that we have gotten along fine with the DDL1 and DDLm core dictionaries not, technically, permitting multiple
instances of those categories to be presented in the same data file at all. In mmCIF, on the other hand, substantially all categories are loopable in principle, with many of them associated together indirectly via the ENTRY category and its _entry.id attribute.
Inasmuch as _entry.id "identifies the data block", however, that amounts to a distinction without much difference.
But mmCIF’s ENTRY category is nevertheless instructive. Formally, many categories defined as Sets in the DDLm core are associated with each other in mmCIF not
by having global nature but by referring to the same ENTRY. This arrangement is similar to what is called a "star schema" in data warehousing: instead of a multitude of individual entities being global (which cannot generally be accommodated in a data warehouse)
or all having direct relationships declared with a large number of other entities, they are instead all related to a single central entity; the relationships can be visualized as emanating in a star-like pattern from that central entity. In such a data warehouse,
the central entity often represents a point in time; it constitutes the dimension along which all the other entities can jointly and concertedly vary.
So suppose we took the ENTRY idea from mmCIF, but allowed a block to contain multiple ENTRYs? As far as I can determine, that’s consistent with the machine-readable
parts of the definitions of ENTRY and _entry.id anyway, though it seems inconsistent with their prose descriptions. In that way, a data file could be valid against mmCIF and nevertheless describe, say, multiple CELLs, without there being any ambiguity about
which CELL went with which REFLNS. That’s similar to what we want to be able to do, but it doesn’t quite get us everywhere we want to go. The problem that we are grappling with can be viewed as how to deal with a situation wherein we want or need a different
pattern of relationships between categories than the one described by the relationships with ENTRY.
James’s proposal #2 approaches the problem from a different angle. It acknowledges that there is more than one possible pattern of categories and relationships
characterizing a data set, and it designates these as "schemas", which is indeed an apt
term. It uses the category label 'Set'
or maybe 'Global' (which I prefer for this purpose) to define a pattern of 1:1 relationships
that serves as a functional substitute for mmCIF’s explicit relationships between ENTRY and other categories; it introduces a mechanism for declaring that a given data file in fact complies with a different schema than the default; and it provides a mechanism
aimed at helping software determine whether and to what extent it can correctly interpret the file’s contents. At that high level, I don’t disagree with any of it, but we’ve gone several rounds over the details. Our main sticking point is related to how
the relationships among categories should be described in dictionaries -- especially those that to date have been implicit in categories being defined as Sets.
Now suppose we combine the high-level idea of providing for multiple schemas with the mmCIF star schema structure. The DDLm core can model each distinct schema
as a simple category and the hub of its own star schema, like mmCIF’s ENTRY. Existing categories can participate in more than one of these where appropriate, though initially there would be only one. Converting the existing DDLm core to this structure would
involve creating one new key in each current Set category (mmCIF already has these keys), and possibly child keys in other categories. It does not necessarily affect existing data files at all, because we can define default values for the various keys. In
this way, all needed keys can be explicitly defined, with a much more modest overall number of keys than if relationships were expressed directly among all categories, and consequently with much less impact when new categories are added.
This also provides a fairly clean way to deal with SPACE_GROUP, and with any future categories that present a similar problem. Whereas with categories such as
CELL we could enforce the restriction of one CELL per hub instance by making CELL’s category key be a child key referencing the hub category, we could reverse that for SPACE_GROUP and any similar category: give the hub category a child key referencing SPACE_GROUP.
To wrap it all together and make it easier for software authors to deal with, we can add
_audit.schema or something like it. One variation that occurs to me would be to have _audit_schema.name and _audit_schema.multiplicity, with the former taking as its values
the names of schema hub categories, and the latter taking values from an enumerated set describing whether that category is present and if so, whether it is restricted to a single value. This would provide a fairly easy mechanism by which data files could
advertise their structure to consumers, and for software to gauge whether they can handle the data.
From: ddlm-group [mailto:email@example.com]
On Behalf Of James Hester
Sent: Monday, June 20, 2016 2:36 AM
To: Group finalising DDLm and associated dictionaries <firstname.lastname@example.org>
Subject: Re: [ddlm-group] Further discussion of proposal #2
To summarise at the top, my principal objection to the 'default key' proposal is that it produces more complex dictionaries (more keys) with interactions that are initially surprising to a casual reader.
I think our goal here is to come up with semantics that can (i) replicate DDL1/DDL2 'global category' behaviour and (ii) allow these global categories to become multi-packeted, with simultaneous loss of 'globality'.
'Global' categories (what I have referred to previously as 'Set' categories) are just a tool for simplification of dictionaries, and so the more complex we make their operation, the less benefit they provide. Likewise, the mainstream behaviour of feature
(i) should be as easy as possible to use.
Proposal #2 as it currently stands (the 'Set' proposal) envisaged that the 'globality' of a category would be removed when using datanames defined within a separate dictionary (mostly key datanames), and software
should use the _audit.schema dataname and potentially _audit.conform to shield itself from the change in meaning that this entails.
The 'default keys' proposal that John has outlined instead envisages making almost all 'Set' categories into 'Loop' categories, defining keys for them, and giving those keys default values. John has suggested
that this does not now involve a change in DDLm, because the semantics of having a default key are clear - the dataname can be left out if there is only one packet. However, a 'global' category with only one packet does *not* (currently) act like a 'Loop'
category with only one packet, because (unlike a single-packet 'Loop' category) the values appearing as non-key datanames in the 'global' category may be assumed when interpreting values from all other datanames in all other loops. 'Global' categories really
are different to 'Loop' categories for this reason, regardless of whether or not a key dataname is provided.
This difference between 'Global' and 'Loop' categories could be removed completely if all of the global category child keys were defined in parallel. In this case, the 'Global' category no longer acts 'Globally' but only in those categories
for which a child key is defined. This 'simplification' comes at the expense of a whole lot of keys - in some categories, a key for every 'Set' category currently defined. At this point we have lost the practical simplification that we had obtained from
'Set' categories to start with. So, either you accept a change in DDLm (additional consequences of a default key) and define the child keys at a future date in another dictionary, or you keep DDLm unchanged and include the child keys in the main dictionary
immediately, throwing out the considerable simplification afforded by having global values. I would be against the latter option as it introduces a bunch of rarely-used key definitions into the main dictionary and is likely to be confusing to a casual programmer.
(We could of course alternatively adopt the blanket rule that values appearing in a single-packet loop act globally with identical 'disappearing key' behaviour. While this is true enough mathematically, it now becomes permissible to drop
keys that have up until now been required even for single packet loops and loops with foreign keys that point to those single-packet loops, and this would break current software. So I exclude this as an option, even if it is an elegant rule.)
So, given that we are stuck with two types of 'Loop' category, I would prefer communicating this clearly up front in the _definition.class tag, rather than relying on the presence of a default key value. What
I think might communicate better than the current 'Set' definition, however, is a change from 'Set' to 'Global' (or 'Overall'), with a definition something like:
A special type of 'Loop' category. When single-valued, (i.e.
key-value pairs or single-row loops) datanames from a 'Global'
category provide overall values for use in interpreting any
other values in a datablock. Global categories may only be
looped where a key has been defined.
I'm not sure if this is more likely to meet with approval.
I have added some more comments in John's email below.
On 18 June 2016 at 08:31, Bollinger, John C <John.Bollinger@stjude.org> wrote:
Dear James and Colleagues,
Comments in line below.
On Thursday, June 16, 2016 9:23 PM, James Hester wrote:
> I'm not at all concerned about tweaking DDLm. The proposed update to DDLm is a clarification and an extension, because the semantic interpretation of existing files would be unchanged. Is there any particular reason you are concerned about such measured
changes to DDLm? From my point of view DDLm is the lowest-impact area of the framework - very few people actually care *how* we express the meaning of a dataname, as long as that meaning doesn't change, and those that do care deeply about DDL in general (in
my experience, databases) have not done any work on DDLm yet.
Perhaps my concerns are misplaced, but it seems to me that the DDLs are the locations of greatest semantic leverage in our framework. On one hand, that means that we can make a large impact with changes there, but on the other hand it means that even small
changes there can have large unintended side effects. Indeed, although I am unaware of any explicit assertion to this effect previously, it seems to me that we should have at least the same commitment to the stability of definitions in our DDL dictionaries
that we do to the stability of definitions in our data dictionaries. But perhaps we can relax that a bit for DDLm, given that its use is still small.
Very little DDLm software has been written, and mostly by those in this group. A lot of thought and negotiation (I believe) has gone into DDLm, so we should not be too cavalier with our changes. Now is the best time to make them rather
than later when we might hope for more widespread adoption.
> I'm not opposed to the concept of a default key value per se, I'm just unclear as to why you are arguing that this needs to be defined in a cif_core 'Set' category as opposed to an add-on dictionary.
I'm arguing that a category that has a key and permits multiple values per item is a de facto Loop, and that it is best to in fact define such a category as a Loop so that that is clear. In that case its key must be expressed in the dictionary that defines
the category. It would also be acceptable to classify such a category with some new label, but in that case I still think it would be most sensible to define the key in the same dictionary that defines the category itself.
See my comments at the top of the email. I have provided a new label and definition, which indicates that the category can be looped, and under what conditions multiple packets may be expected. Perhaps this is acceptable?
I'm also arguing against the "magic keys" aspect of Proposal #2. I don't like magic, a.k.a. special cases, in specifications or in software, and I have presented a viable alternative in the form of default key values.
The reason for the special case 'Set' category is the considerable simplification it offers. We trade complexity of behaviour in one place for simplicity elsewhere. And we are ultimately stuck with it because of DDL1.
I'm furthermore arguing that even if we do give keys to Sets, wherever a category key or child key is itself defined is the proper place for any applicable default value for that key to be defined. The default value is an attribute of the definition of the
key item, so I see only negatives to physically separating the two.
Absolutely, I wouldn't argue with this.
>> Let's consider the SPACE_GROUP category, since it sparked this whole discussion. I append a cut at what I think we should do with it (only frames containing modifications are presented); I think I have marked all the changes and additions within via CIF
comments. I rarely wrangle dictionaries, so I apologize for any errors I have committed. The key defaulting presented within formalizes how, when, and why SPACE_GROUP's category key and the associated child key in SPACE_GROUP_SYMOP can be omitted from data
files. To the best of my knowledge, nothing within relies on any DDLm changes.
> I think I understand your proposal to be using the existence of a default key value to signal that the key may be omitted in a single-value loop, *and* that child key datanames in other loops that would otherwise contain them may be omitted in this case.
I guess you can describe it as a "signal". I view it as deeper and more organic: where an explicit parent or child key may be omitted from data files, that is a direct consequence of the fact that it has a default value. That dictionary-driven software should
handle such omissions naturally is also a consequence. These items can be omitted because they still take well-defined and suitable (default) values in that case.
I don't think I'm suggesting any change to the defined meaning of _enumeration.default; I'm just applying its existing meaning to the problem at hand in a way that we have not done before. The significance pertains not to _enumeration.default itself, but to
its combination with a category key. That's not a change, it's a discovery. Even so, the underlying idea is not actually new. One can view it as a specific case of the same thing expressed by DDL2's _item.mandatory_code taking the value 'implicit'.
See my comments at the beginning for why I think there is more than just logical consequences going on here i.e. there is global behaviour.
> I'm not clear whether you propose that these changes should happen in cif_core, or in an add-on dictionary.
For space_group, the dictionary changes should be applied to the core, in order to make the DDLm core consistent with our other dictionaries. I am generally inclined to put future (re-)keyings of core categories directly into the core dictionary as well, but
that's a weaker opinion. Furthermore, I think there may be a way to do this so that we avoid an explosion of child keys, but I haven't worked all the way through that yet.
Your proposal on child keys would be interesting as I argue above that an explosion of child keys is a drawback and essentially removes the advantage gained by having global categories.
> In any case, I agree that this can be made precisely semantically equivalent to the 'Set' proposal, due to the fact that a default key value makes no sense in general and so the meaning of a default value for a key may be overloaded as you have done, with
no implications elsewhere. This is still a change to DDLm, because the presence of _enumeration_default in certain definitions now has new implications (not that I'm opposed in principle to changing DDLm).
I agree that the "magic keys" aspect of Proposal #2 and the default keys approach I have presented both enable categories to have keys that are not expressed explicitly in data files. The former does it by fiat; the latter does it in a manner consistent with
DDLm's existing semantics, even if our dictionaries have not exercised DDLm in quite that way before.
I agree that a default key value is not necessarily sensical for every present or conceivable category, but I disagree that I am overloading any definition, or that I am proposing a change to DDLm.
I would see no problem in a separate dictionary defining the key and default value for a 'Global' category as I've defined above. Essentially, conformance to this separate dictionary erases the 'Global' nature of the category and turns
it into a normal 'Loop' category with default key, so that datafiles created according to the original specification remain valid with the new dictionary - we have in fact elegantly expanded the ontology.
Default key values do not make sense for categories that rely on natural keys, as does mmCIF's atom_type category, for example. Atom_type's key, _atom_type.symbol, is the chemical symbol for the element whose characteristics are described; it is a natural
key because it has significance beyond distinguishing one atom_type from another. In other words, it is not just a key, but also part of the data.
On the other hand, space_group does not use a natural key, but rather a surrogate key -- one whose values have no inherent meaning other than to distinguish between different space_groups presented in the same data file. If only one space_group is presented
then any key for it will do, because the keys are arbitrary. A default value for such a key is perfectly sensible.
Now, consider this: what kind of key will any new category have if that category requires one or more existing Set categories to become looped? We have previously discussed possibilities such as twin_component and variant, but as far as I can tell, these do
not afford any clear, non-trivial, natural keys. Addition of any category that relies on a natural key would require existing sets to be looped only if that category's key is inherently single-valued with respect to those sets. I'm in fact having trouble
seeing the circumstances under which it would make sense to add a new category that has a natural key and that requires existing sets to be looped. But even if we did discover a new category with a non-trivial, natural, candidate key, we always have the option
of choosing a surrogate key instead. Indeed, that's what was done with space_group -- _space_group.name_Hall is a candidate key, I think, but we chose a surrogate key instead. If we choose surrogate keys then default key values present no semantic problem.
I agree with this - I'm not arguing that default key values are somehow bad or present problems, only that the 'global' behaviour is not captured.
> My preference would still be for the 'Set' proposal, because the semantics are wrapped up in a single enumerated value, at category level, rather than arising from an interaction between attributes of a particular dataname inside that category. I do not
see any other distinguishing features. I believe that for programmers, dictionary authors, and casual dictionary readers, the 'Set' proposal is more accessible, as the particular special behaviour of the category is flagged explicitly and concisely, in the
category definition, and described in a single place in the DDLm attribute dictionary.
Moreover, even if we did provide magic key behavior for Sets, I am not convinced that all the constituencies named would necessarily consider that a win, because it weakens the concept of a Set. There is a tremendous difference between "the items in a Set
category take only one value each" and " the items in a Set category *ordinarily* take only one value each", especially when "ordinarily" really means when the data describe a particular kind of thing to which we have ascribed special status. In many respects,
programming for, using, or interpreting the latter (the magic keys version) are all more difficult than programming for, using, or interpreting the former (the current version).
OK, point taken, I did say my objection wasn't critical.
> You will notice there is semantic convenience in referring to a category as a 'Set' category, rather than 'a category that has a default key value defined'. If you propose changing the cif_core dictionary rather than using an add-on dictionary, then the
'Set' proposal involves zero changes, whereas the default_value proposal involves a single extra key definition and adjustment to the definitions for each 'Set' category. Both these objections are not particularly critical, of course.
The semantic convenience described comes at the cost of weakening the concept of a 'Set', and as a result, the comparison presented involves inequivalent expressions. The magic keys analog of 'a category that has a default key value defined' is 'a Set that
has a category key defined'; these don't seem very different in weight to me. If we suppose that a Set may have one or more keys defined in a different dictionary than the one in which the Set itself is defined, then additionally we may not even be certain
which kind of Set we're talking about, and if that ever changes then we cannot be confident of being able to recognize that from the dictionary at hand. That is of course where _audit.schema and audit_conform come in, but I am not much liking the idea that
an applicable item definition, taken in context of its dictionary, may not completely define the given item.
That is where we started - if we are to allow datanames to used with global meaning and in multi-packet loops, then we are talking about different meanings, and only something like _audit.schema can insulate
software from that. If we are to exclude changes in meaning, we have to define all child keys up front and then we need _audit.schema even more than before, as _audit.conform won't help. In the 'all child keys defined up front' scenario, we completely abandon
global categories and _audit.schema becomes the signal as to when a datablock can be interpreted as for the old cif_core.
> Ultimately, this is going to be a matter of taste as the semantics can be made identical, and so I don't know quite what else you or I can say to convince each other on this point. We may have to rely on
our colleagues to decide.
We do seem to have both settled into our positions. Would it sway you at all if I successfully devised a solution to the child key proliferation problem? I have some ideas in that direction that I haven't fleshed out yet.
It could indeed sway me as I think this is at the core of my objection. If we could effectively define all the child keys, while at the same time keeping the key definitions from swamping out the meat of the dictionary, and allow for the
appearance of future 'used to be global' categories like twinning and variants adding their own child keys, then it would be worth serious thought. I'm pretty sure dREL can be brought along with whatever variation you propose.
>> Note, by the way, that I think the particular changes presented, or something very like them, are needed regardless of what we choose for the general case, because the DDL1 core and mmCIF are already structured this way.
> I was perhaps too diplomatic or long-winded in previous messages. The incorporation of space_group into cif_core as a looped category was a mistake that we must *not* perpetuate. We either correct it by dropping it from DDLm cif_core, which is impossible
due to widespread DDL1 usage (as a 'Set' category), or we fix the semantics. So, in the case of space_group we can feel ourselves bound only by widespread current usage, not by the contradictory semantics of the DDL1 version.
I accept that the deprecation of SYMMETRY and SYMMETRY_EQUIV in favor of SPACE_GROUP and SPACE_GROUP_SYMOP was a mistake, but whatever fix we contemplate should adhere to our policy of keeping definitions stable, at least as well as we are able to make it do.
Moreover, how to deal with SPACE_GROUP is a somewhat separate issue, because it involves definitions that already exist, as opposed to definitions that we may write in the future. It makes for a reasonable test case for our future direction, but it may be
that a different solution is more suitable here than whatever we decide to do in the future, when we have no legacy definitions to deal with.
Our policy of keeping definitions stable is not an end in itself, but a logical requirement born of the need to guarantee that software that is already written remains valid. If everybody is using unlooped SPACE_GROUP to read and write
structures I don't see any issue in fiddling with the meaning, as long as any changes are consistent with that expectation of an unlooped value.
I'll have more to say about these particular cases, in a separate message.
T +61 (02) 9717 9907
F +61 (02) 9717 3145
M +61 (04) 0249 4148
Email Disclaimer: www.stjude.org/emaildisclaimer
Consultation Disclaimer: www.stjude.org/consultationdisclaimer