Sunday, October 18, 2015

Versioning Domain Entities with Ordinary SQL

In any system with mutable entities saved in an object store, it's inevitable that you'll want to audit the changes made to a particular entity. Explicitly versioning your entities solves this problem, and it's often pretty useful to end users of a sophisticated application, say for undo purposes. However, it's not always so clear how to transform a mutable entity into an versioned entity.

There are also a number of complications to versioning mutable objects:

  1. Handling parent-child relations such that a child change does not propagate to every parent
  2. Supporting efficient queries on history
  3. Supporting efficient queries for the newest version
  4. Space complexity of data representation
  5. Concurrent changes

Some of the above properties also have a natural tension, ie. supporting efficient queries often require storing more data so their space complexity is worse. I'm going to provide an overview here of a very simple schema for versioned entities with the following properties:

  1. Queries on history are just as efficient as queries on current data
  2. Queries on history have the same form as queries on current data, ie. no separate history table that must be handled specially -- in fact, you can use the exact same queries for both history and current data, thus eliminating all query duplication!
  3. Parents and children can evolve independently, ie. data update is "strictly disjoint access parallel" in transactional concurrency parlance: as long as two updates don't update the same object, they will both commit successfully

Schema

The schema changes required are quite straightforward for any entity with a primary key (which should be all of them!):

  1. Add a version number column of type int or long: this is a monotonically increasing version number, where version=0 is used when the object is first created, version=1 is after the first change is committed, etc.
  2. Ensure the PK index orders the version number in DESCENDING order, because the highest version number is the most current version
  3. Change the primary key to be composite over the existing PK and the new version number -- the original PK is now called the "logical id"
  4. Add a revisionDate column storing the date the version was committed
  5. Add an index over the logical id and the revisionDate, again with revisionDate in DESCENDING order

That's it! The original entity's PK is now a logical id rather than an actual instance id, so accessing a particular version of an entity now requires both the logical id and the version to access. The latest version is simply the entry with the largest version number.

Queries

Supposing we have an entity Foo whose PK is called "id", and :columnName is the syntax for a named query parameter. Ordinary queries on entities now become the following queries on versioned entities:

  • Insert a new entity:
    INSERT INTO Foo(id, ..., version, revDate)
    VALUES (:id, ..., 0, GETDATE())
  • Get the most current version:
    SELECT TOP 1 * FROM Foo
    WHERE id=:id
    ORDER BY version DESC
  • Get a version before a specific date/time:
    SELECT TOP 1 * FROM Foo
    WHERE id=:id AND revisionDate <= :rev
    ORDER BY version DESC
  • Update an entity:
    INSERT INTO Foo(id, ..., version, revDate)
    VALUES (:id, ..., :cur_version_plus_1, GETDATE())

Updates on entities are now inserts, so the entities at the data store level are now effectively immutable and the entity table is an append-only log of all changes to those entity types. Since the indexes are ordered so that the latest version is always first, the impact on accessing the most current version should be minimal. Accessing older versions are similarly easy, consisting of a query on the revisionDate column.

We use the revisionDate and not the version number when querying over history to easily support queries over more than one entity. Since each entity evolves its version independent of all others, the version numbers of two connected entities, say Foo[id=1] and Foo[id=2], are completely unrelated. The former might have version=999, and the latter version=1. The change that caused Foo[id=1] to increment its version from 998 to 999, could have also been the one to change Foo[id=2] from version 0 to 1. So if you look at Foo[id=1] at version=998, the only way to see the correct version of Foo[id=2] is to return the first version whose revisionDate is <= the Foo[id=1].revisionDate for version=998.

Fortunately, this makes the queries extremely simple, since a complicated query for a whole set of interconnected entities needs only a single date parameter. In other words, you can compose two independent queries using the same revisionDate parameter name via simple concatenation into a multiquery, and it just works.

Furthermore, your UI then only needs to pass around an optional revisionDate parameter to switch between queries on the latest data and queries on history. In fact, switching the query isn't even needed, because you can use the same query for both!

SELECT TOP 1 * FROM Foo
WHERE id=:id AND (:rev IS NULL OR revisionDate <= :rev)
ORDER BY version DESC

As promised, this query unifies the queries for the most current version, and any historical version of an entity. If the :rev parameter is set to NULL, then the most recent version is returned. The query optimizer can easily eliminate the redundant WHERE clause, so there should be no performance impact. If :rev is present, the redundant NULL check should be eliminated by the query optimizer and the last change before the given date/time would be returned. Alternately, you can simply always use the query with the revisionDate and simply pass in DateTime.Now, which will always return the latest result.

Concurrency

Being strictly disjoint access parallel, this schema on mutable entities is pretty much as good as it gets concurrency-wise. The entity table is essentially an append-only log of changes, and the new composite primary key ensures that any entity can undergo only one change event at a time. This ensures that any invariants in the application's domain model are preserved once committed to the entity store.

This is also the ideal optimistic concurrency model for any CRUD application.

Space Complexity

The space complexity is obviously sub-optimal since every version is stored. However, managing this space is a breeze, since you can easily write a query to delete all entity history before the current version, or keep the N latest versions, or basically any storage policy that makes sense for your business.

One serious concern are "large" columns in an entity, say an unbounded text column, or binary data. In this case, you should perform an additional refactoring of your schema to model git's way of doing things: add a Blob table that holds the data indexed by its hash, then just store the hash in the entity table. This keeps it compact, and ensures you only ever store one version of a binary blob, which you should probably be doing anyway.

Variations

There are a number of variations on this schema, like normalizing the revision date into a separate Revisions table to which you can add a userId to blame for the change, etc. But then the Revision table becomes a central bottleneck to insert performance and updates are no longer strictly disjoint access parallel. Furthermore, all select queries must now join on this Revisions table to find the revision date. I think the schema I've laid out provides the optimal performance tradeoff with little enough space overhead.

Versioned ORM

Standardizing this schema as the default operating mode for an ORM would be ideal. This sort of ubiquitous schema management should be automatic, and now that ORMs are moving to code-first designs where they manage the whole schema anyway, this should be a snap right?

The API interacting with the ORM remains largely unchanged, but all change submissions transparently become inserts, and all select queries automatically have the universal query form described above. A query on history would then simply be a query that's started with a date or version parameter, ie. instead of context.Foo.Where(...), it would be something like context.Foo.Before(DateTime?).Where(...).

Summary

I'm not the first to describe this schema for versioning, but I haven't seen a description quite so comprehensive, detailing the queries and its concurrency properties.

These days most of the buzz is around CQRS and event sourcing. This schema has some connections to event sourcing, since entity tables are now append-only logs with the entire history remaining intact. However, the schema of the entity change events aren't domain events as they are in event sourcing.

CRUD applications don't get much attention anymore, but I agree with Martin Fowler that most simple domains are still better served by a CRUD model than by CQRS, at least given our current tools. Domain entities usually define proper concurrency boundaries (and if they don't, your domain model is probably wrong!), so this schema is a natural fit that still ensures a domain model can naively enforce its invariants in ordinary code, and if those changes succeed then those invariants are preserved in the entity store.

Sunday, October 4, 2015

Efficient Curried Application

I just wanted to jot down some thoughts on compiling eval/apply curried application using an efficient register-based calling convention. This translation is actually quite simple once you get the trick. Leroy's Zinc abstract machine showed the way: evaluate arguments right-to-left instead of the typical left-to-right. The ZAM wrote these arguments to a stack, but in our case we want them to go to the register file first.

So we evaluate arguments right-to-left and write them into registers left-to-right, ie. argN goes in reg0, argN-1 goes to reg 1, ... arg0 goes to regN. At some threshold, we'll run out of registers. If your program is in direct-style with a stack, this threshold is often defined by the callee-save registers of the native calling convention. If your program is in CPS form, then you can use the entire register file.

Once you run out of registers, you have to spill the remaining arguments to the stack. However, the stack is simply a register pointing to some base address from which all other values are accessed by an offset from that base. But so is the closure record! If you lay out your closure record in memory so it has the same layout as the stack frame would, then you can just treat it like a small stack. This can save many memory access operations moving data between closures and registers.

Suppose our register spill threshold is 2 registers (reg0 and reg1), with a third register left for the remaining arguments (reg2), and consider the following C-style function:

int foo(int x, void *y, int z);

Here we have three arguments, one of which will need to be spilled to the stack since we can only allocate two when calling foo. Arguments z and y will go in registers, while argument x will go on the stack. However, the compilation of foo will require a certain layout in memory for optimal efficiency as well (assuming 32-bit pointers):

foo_papp2:
  loadi reg0, reg2 // reg0 = *reg2
  sub1 reg2, 4     // reg2 += sizeof(int)
foo_papp1:
  loadi reg1, reg2 // reg1 = *reg2
  sub1 reg2, 4     // reg2 += sizeof(void*)
foo:
  ...              // foo: reg0=arg2, reg1=arg1, reg2=&struct{arg0}

Hopefully the pseudo-assembly is clear. Function foo gets "prologues", one per argument which is used for partial applications. Each prologue loads an argument from the "stack"/closure record into the registers and decrements the pointer so it points to the next argument. By the time all the registers are used up, you should have exactly the "stack frame" that you need left, and the registers-passed arguments are all in their proper place.

For instance, when you partially apply foo to 1 argument, the function pointer value of the closure record is actually pointing to foo_papp1, and when you partially apply with two arguments, it's pointing to foo_papp2. This pattern is the same for any number of arguments, and the code will all have the same size for each argument, which we will need later when building partially applied closures.

Now consider the following curried applications of foo:

int(void*,int) foo_1 = foo(0);
int(int) foo_2 = foo(0, somePointer);

Hopefully the pseudo C-style closure syntax is clear, ie. int(char) is a closure taking a char and returning an int. The above are two different partial applications, foo_1 which applies only the first argument, and foo_2 which partially applies the first and second arguments. These are what the closures will look like in memory:

foo_1 →
function = &foo_papp1
arity = 2
applied = 1
x = 0
foo_2 →
function = &foo_papp2
arity = 1
applied = 2
y = somePointer
x = 0

Now when you try to apply values foo_1 and foo_2 to the remaining arguments, this actually performs a dynamic check to see how many arguments are required. For instance, fully applying a closure with the signature int(int, int) elaborates to the following pseudo-code:

obj_t returnValue;
RETRY:
switch (closure→arity) {
case 0:
  closure = closure→function();
  goto RETRY;
case 2:
  returnValue = closure→function(arg0, arg1);
  // reg0=arg1, reg1=arg0, reg2=&closure->arity + 1
  break;
case 1:
  closure = closure→function(arg0);
  returnValue = closure→function(arg1);
  break;
}

You have to dynamically check if you're applying all remaining arguments. I handle the case of arity=0 to show that intermediate closures do exist, but you don't typically need that specific case.

When you're performing a partial application, you also have to check the arity and build intermediate closures referencing the proper function address, which changes depending on the arity. Given a closure and performing a partial application, you have to make adjustments to point to the new function pointer. So given a 3 argument closure to which you're applying n < 3 arguments:

closure nclo;
obj_t returnValue;
if (closure→arity == closure→applied + n) {
  returnValue = closure->function(arg0, ..., argn);
} else {
  nclo = malloc(2 * sizeof(void*) + closure→arity - closure→applied + n * sizeof(void*));
  nclo→arity = closure→arity - n;
  nclo→applied = closure→applied + n;
  nclo→function = closure→function - SIZE_OF_FUNCTION_PROLOGUE;  // compute offset of new entry point
  nclo→arg0 = arg0;
  ...
  nclo→argn = argn;
  memcpy(&nclo->argn + sizeof(void*), &closure→applied + sizeof(void*), closure→applied * sizeof(void*));
}

The most general case to handle is when you're in fully polymorphic code for which you cannot know if applying the closure you have is a partial application. This is a combination of the last two elaborations, but still requires only a simple switch statement to handle all cases.

Now the most boring case: when a function is not a closure and it's being fully applied with all arguments at a call site: well in this case you fill the registers as usual, then you push the remaining arguments on the real program stack, and pass the pointer to that activation frame as the "stack" argument. Simple!

This is only a high-level overview for myself, so no doubt I've skipped some important details, or gotten them wrong while jotting these notes down. Please let me know if you spot something!

Summary

Hopefully that didn't ramble so much that the core idea was obscured. The core idea is that all closures are functions taking a fixed number of arguments defined by the number of caller-save registers. These arguments are passed in registers, and a last register is reserved for a "stack frame" holding the remaining arguments. The closure record can act as this stack frame, so you don't have to copy values between records and the stack on closure application.

Furthermore, each function has a "prologue" generated for each argument laid out prior to the actual function being executed in memory. These prologues are used when applying partially applied functions, and they load any previously applied arguments from the closure record into registers, assuming all the reserved registers aren't already used.

This is basically an implementation of the eval/apply strategy, with a few tweaks to reduce the amount of redundant code generated and reuse as much memory as possible. I don't think there's anything particularly novel, but I haven't seen all the details laid out like this (although I may of course have ommitted some details I consider "obvious", even if they aren't to others). If you have any further suggestions, I'd love to hear it!

Limitations

I don't discuss any of the complexities surrounding unboxed types of non-pointer size, which considerably complicates calling conventions. Most functional languages require a universal representation for this reason, usually of pointer size, which is what I assume here. See my series on Garbage Collection Representations to get a feel for how these values translate to machine code, and how this would impact closure representations.