[hibernate-dev] 6.0 - Type system

[Steve Ebersole]

Right, and that exactly lines up with what I am proposing.

If the intent of "customize" is to describe new Java types (e.g. Java 8
temporals prior to our explicit support) the tht is the role of a
JavaTypeDescriptor, specifically a BasicJavaDescriptor.  They would
register a BasicJavaDescriptor describing the type.

If the intent is to model a non-supported SQL type then that would mean
adding a new SqlTypeDescriptor describing that type, although that will
often also mean adding a new BasicJavaDescriptor describing the Java
mapping of that SQL type.


On Mon, Jan 23, 2017 at 7:00 AM Vlad Mihalcea <mihalcea.vlad at gmail.com>
wrote:

Hi,

Related to your questions:

he main thing I wonder about is what we mean by "custom
types" in terms of what exactly is being customized?  And how does that
relate specifically to BasicType versus EmbeddedType versus ...?


Most of the time, the users want to take advantage of various database
types that are not universally supported by all RDBMS: JSON, Money (SQL
Server).

On the Java side, I don't see what we can customize because we already
provide all the basic types, and for everything else, users can compose
those into Embeddables. The Java 1.8 Date/Time are an example of what users
would like to customize in case we didn't support this already. But even if
Java 1.9 adds other basic types, chances are that we are going to support
them natively, meaning that users will still not need to add a custom Type.

So, I don't see how a Hibernate user will customize the way Embeddables,
Enums, Entities or Collections are being stored or loaded from the
database. The exception to the rule is a recent Pull Request from someone
who wants to support PostgreSQL arrays. But this falls back into the same
category as before: database types that are not universally supported by
all RDBMS.

Vlad

On Thu, Jan 19, 2017 at 9:55 PM, Vlad Mihalcea <mihalcea.vlad at gmail.com>
wrote:

There's a lot to dig in here. I'll have to get the branch and study the
changes, to come back with some opinions.

Vlad

On Thu, Jan 19, 2017 at 7:34 PM, Steve Ebersole <steve at hibernate.org> wrote:

We are getting pretty far along on the 6.0 changes and I wanted to start
a(nother) discussion about Types in 6.0 to get feedback and thoughts on a
few topics.

First a quick break down of JavaTypeDescriptors, SqlTypeDescriptors, Types
and "persisters"...

(a lot of this is the same from pre-6.0, just making things more explicit)

JavaTypeDescriptors and SqlTypeDescriptors are the "lowest level", so let's
start there.  A JavaTypeDescriptor is a descriptor of a given Java type.
That is, it provides Hibernate with information about the Java type.  Is it
a numeric type?  How do we compare 2 values of this type?  How do we make a
deep copy of a value of this type?  Etc.  SqlTypeDescriptor is the same,
but for a database type (VARCHAR, BLOB, etc).  These 2 work together to
perform reading and writing at the JDBC-level.

We decided to broadly categorize JavaTypeDescriptors based on the JPA type
categorizations:

   1. BASIC - BasicJavaDescriptor
      1. TemporalJavaDescriptor
      2. NumericJavaDescriptor
   2. MANAGED - ManagedJavaDescriptor
      1. EMBEDDABLE - EmbeddableJavaDescriptor
      2. IDENTIFIABLE - IdentifiableJavaDescriptor
         1. MAPPED_SUPERCLASS - MappedSupercassJavaDescriptor
         2. ENTITY - EntityJavaDescriptor


Type (org.hibernate.type.spi.Type) represents a combination of a
JavaTypeDescriptor and one or more SqlTypeDescriptors in relation to a
specific "non-root Navigable domain value".  Navigable is a query-focused
contract (SQM/HQL/Criteria) so I wont get too deep into that here.  At a
high-level t is similar to JPA's Bindable except that it applies to
Collection indices (or map-keys) and elements (or map-values) as well.
Navigable essentially represents an named navigation one can perform in a
query.  The root Navigable is always an entity (EntityPersister).
EntityPersister is the only Navigable that does not expose a Type.  (There
is an EntityType, but it represents entity-valued non-root Navigables such
as a ManyToOne).  All other navigables expose a Type.  That is all a
long-winded way to say that Types represents that Java/SqlTypeDescriptors
for a role-based Navigable.

Like the categorization discussed above for JavaTypeDescriptor, Type has a
similar categorization:

   1. Type
      1. BasicType
         1. TemporalType
      2. AnyType
      3. ManagedType
         1. EmbeddedType
         2. IdentifiableType
            1. MappedSuperclassType
            2. EntityType

It is important to keep in mind that these represents a specific reference
to thse things in regards to a Navigable.  E.g. an EntityType is the "type"
of a SingularPersistentAttribute that is a ManyToOne - it points to the
corresponding EntityPersister but it also represents the FK columns to
refer to the entity.  It is a role-based Navigable.

Historically reads and writes have all routed through the Type (with
certain Types delegating much of that to persisters).  That will no longer
be the case in 6.0 as one of the main design goals for 6.0 is to re-write
how Hibernate reads and writes (mainly reads) values from JDBC.  The major
shift here is to read all values from JDBC using a "SqlSelectionReader"
equivalent to a BasicType.  These values are read and held in an array that
"readers" then know how to access (positionally) and use.  Most of that
design is beyond the discussion here, but it useful to understand.  It is
discussed in the design.adoc in my orm-sqm poc repo for those curious.
Long story, short... Types no longer directly implement JDBC read/write
which we will come back to later.

PersistentAttribute and the other Navigables now take a role in JDBC
reads/writes.  AttributeConverters and other read/write-related concerns
have been moved to these contracts.  Again, most of this is covered in the
mentioned design doc.

Since Type no longer directly implements JDBC read/write operations I think
it is important to ask ourselves what exactly we see as "customizable" wrt
each Type.  Is that different for each category, or the same across all
Type categories?  E.g. I know of no customization of EntityType as it
exists in 5.x, and tbh I am not even sure what that would mean.  BasicType
obviously has some parts that we want to allow users to override, but is
that really best achieved through a custom BasicType?  Or is it better
served by allowing custom JavaTypeDescriptor/SqlTypeDescriptor and/or
SqlSelectionReader?  What about EmbeddedType?  CollectionType?  This would
affect @TypeDef and Type registration methods specific to customizations.

Persisters for the most part continue to serve the same role they have in
the past with a few additions and some changes...

One addition was the creation of an EmbeddedPersister.  *Embedded*.  This,
like CollectionPersister, models a "role" e.g. "Person.name" as opposed to
the Embeddable Name.class.  Note however that JPA calls it an
EmbeddableType and expects info about the Embeddable (the Class).
EmbeddedPersister is role-based (Embedded) instead, which is a mismatch.
In the case there are more than 1 usage of the Embeddable in different
Embedded roles then we have to decide which EmbeddedPersister to return.
It affects the sub-Attributes information.  We could just return "one of
them" and deal with it for Alpha1, but we should answer what we want to do
there long term.

Collectively, these persisters now implement the JPA ManagedType model
directly.  Another addition was the creation of ManagedTypeImplementor,
IdentifiableTypeImplementor and MappedSuperclassTypeImplementor in the
persister hierarchy.  Which means we can now directly return them in our
JPA Metamodel impl.

That also means implementing JPA's notion of Attributes.  I also needed
something similar for SQM's Navigable contract.  Plus I have been working
towards changing how Hibernate understands Attributes internally
(encapsulation - OO ftw!) for some time anyway, so this all dove-tailed
well.

There are some things we should discuss too in terms of user impact.  We
know up front that we need to move to reading values from JDBC ResultSets
positionally, as opposed to nominally which is how it was exposed in
Hibernate prior to 6.0.  So we know already we will be asking implementors
and consumers of those contracts to make changes.  Given that, we have
*some* liberty in changing these contracts some more.  We just want to be
cognizant of (a) how much we change, (b) who it affects (use cases) and (c)
whether there are alternatives.  For any use cases we determine to be
"valid" use cases, I think we need to make certain that there is some way
to handle that in the new alternatives.

One use case, e.g., is setting Query parameters and being able to specify
its Type.  To a degree we want to be able to continue to support that.  But
I think we limit it to just references to org.hibernate.type.Type (though
"gutted") specifically and remove all others; and temporarily have the new
org.hibernate.type.spi.Type interface extend the old.  This would allow
them to continue to get these org.hibernate.type.Type references in some
fashion and use them as Query parameter type hints.  But I think we should
look at implementing some other parameter type "hints" like accepting
PersistentAttribute/Navigable references, JPA (static) metamodel
references, etc.  These are better, as they would include things like
AttributeConverter whereas the Type reference would not.

Sorry this got so long.  I've had a lot floating around in my head the last
few days as I have worked on 6.0 and I wanted to bring them up for
discussion.  The main thing I wonder about is what we mean by "custom
types" in terms of what exactly is being customized?  And how does that
relate specifically to BasicType versus EmbeddedType versus ...?
_______________________________________________
hibernate-dev mailing list
hibernate-dev at lists.jboss.org
https://lists.jboss.org/mailman/listinfo/hibernate-dev