* Part 1: define ValueTypeSet.
We're going to use this instead of `HashSet<ValueType>` so that we can clearly express
the empty set and the set of all types, and also to encapsulate a switch to `EnumSet`."
* Part 2: use ValueTypeSet.
* Part 3: fix type expansion.
* Part 4: add a test for type extraction from nested `or`.
* Review comments.
* Review comments: simplify ValueTypeSet.
* Pre: put query parts in alphabetical order.
* Pre: rename 'input' to 'query' in translate tests.
* Part 1: parse :limit.
* Part 2: validate and escape variable parameters in SQL.
* Part 3: algebrize and translate limits.
This is for two reasons.
Firstly, we need to track the types of inputs, their values, and also
the input variables; adding a struct gives us a little more clarity.
Secondly, when we come to implement prepared statements, we'll be
algebrizing queries without having the values available. We'll be able
to do a better job of algebrizing, and also do more validating, if we
allow callers to specify the types of variables in advance, even if the
values aren't known.
This adds an `:order` keyword to `:find`.
If present, the results of the query will be an ordered set, rather than
an unordered set; rows will appear in an ordered defined by each
`:order` entry.
Each can be one of three things:
- A var, `?x`, meaning "order by ?x ascending".
- A pair, `(asc ?x)`, meaning "order by ?x ascending".
- A pair, `(desc ?x)`, meaning "order by ?x descending".
Values will be ordered in this sequence for asc, and in reverse for desc:
1. Entity IDs, in ascending numerical order.
2. Booleans, false then true.
3. Timestamps, in ascending numerical order.
4. Longs and doubles, intermixed, in ascending numerical order.
5. Strings, in ascending lexicographic order.
6. Keywords, in ascending lexicographic order, considering the entire
ns/name pair as a single string separated by '/'.
Subcommits:
Pre: make bound_value public.
Pre: generalize ErrorKind::UnboundVariable for use in order.
Part 1: parse (direction, var) pairs.
Part 2: parse :order clause into FindQuery.
Part 3: include order variables in algebrized query.
We add order variables to :with, so we can reuse its type tag projection
logic, and so that we can phrase ordering in terms of variables rather
than datoms columns.
Part 4: produce SQL for order clauses.
* Pre: refactor projector code.
* Part 1: maintain 'with' variables in AlgebrizedQuery.
* Part 2: include necessary 'with' variables in SQL projection list.
The test produces projection elements for `:with`, even though there are
no aggregates in the query. This test will need to be adjusted when we
optimize this away!
This commit turns complex `or` -- `or`s in which not all variables are
unified, or in which not all arms are the same shape -- into a
computed table.
We do this by building a template CC that shares some state with the
destination CC, applying each arm of the `or` to a copy of the template
as if it were a standalone query, then building a projection list and
creating a `ComputedTable::Union`. This is pushed into the destination
CC's `computed_tables` list.
Finally, the variables projected from the UNION are bound in the
destination CC, so that unification occurs, and projection of the
outermost query can use bindings established by the `or-join`.
This commit includes projection of type codes from heterogeneous `UNION`
arms: we compute a list of variables for which a definite type is
unknown in at least one arm, and force all arms to project either a type
tag column or a fixed type. It's important that each branch of a UNION
project the same columns in the same order, hence the projection of
fixed values.
The translator is similarly extended to project the type tag column name
or the known value_type_tag to support this.
Review comment: clarify union type extraction.
This commit:
- Defines a new kind of column, distinct from the eavt columns in
`DatomsColumn`, to model the rows projected from subqueries. These
always name one of two things: a variable, or a variable's type tag.
Naturally the two cases are thus `Variable` and `VariableTypeTag`.
These are cheap to clone, given that `Variable` is an `Rc<String>`.
- Defines `Column` as a wrapper around `DatomsColumn` and
`VariableColumn`. Everywhere we used to use `DatomsColumn` we now
allow `Column`: particularly in constraints and projections.
- Broadens the definition of a table list in the intermediate
"query-sql" representation to include a SQL UNION. A UNION is
represented as a list of queries and an alias.
- Implements translation from a `ComputedTable` to the query-sql
representation. In this commit we only project vars, not type tags.
Review comment: discuss bind_column_to_var for ValueTypeTag.
Review comment: implement From<Vec<T>> for ConsumableVec<T>.
Complex `or`s are translated to SQL as a subquery -- in particular, a
subquery that's a UNION. Conceptually, that subquery is a computed
table: `all_datoms` and `datoms` yield rows of e/a/v/tx, and each
computed table yields rows of variable bindings.
The table itself is a type, `ComputedTable`. Its `Union` case contains
everything a subquery needs: a `ConjoiningClauses` and a projection
list, which together allow us to build a SQL subquery, and a list of
variables that need type code extraction. (This is discussed further in
a later commit.)
Naturally we also need a way to refer to columns in a computed table.
We model this by a new enum case in `DatomsTable`, `Computed`, which
maintains an integer value that uniquely identifies a computed table.
When we started expanding and narrowing type sets, it became impossible
to conclusively know during pattern application whether a type was
known. We now figure that out at the end: if a variable has only a
single known type, we don't need to extract its type tag.
Part 1, core: use Rc for String and Keyword.
Part 2, query: use Rc for Variable.
Part 3, sql: use Rc for args in SQLiteQueryBuilder.
Part 4, query-algebrizer: use Rc.
Part 5, db: use Rc.
Part 6, query-parser: use Rc.
Part 7, query-projector: use Rc.
Part 8, query-translator: use Rc.
Part 9, top level: use Rc.
Part 10: intern Ident and IdentOrKeyword.
mod.rs defines the module and ConjoiningClauses itself, complete with
methods to record facts and ask it questions.
pattern.rs, predicate.rs, resolve.rs, and or.rs include particular
functionality around accumulating certain kinds of patterns.
Only `or.rs` includes significant new code; the rest is just split.
* Pre: Don't retract :db/ident in test.
Datomic (and eventually Mentat) don't allow to retract :db/ident in
this way, so this runs afoul of future work to support mutating
metadata.
* Pre: s/VALUETYPE/VALUE_TYPE/.
This is consistent with the capitalization (which is "valueType") and
the other identifier.
* Pre: Remove some single quotes from error output.
* Part 1: Make materialized views be uniform [e a v value_type_tag].
This looks ahead to a time when we could support arbitrary
user-defined materialized views. For now, the "idents" materialized
view is those datoms of the form [e :db/ident :namespaced/keyword] and
the "schema" materialized view is those datoms of the form [e a v]
where a is in a particular set of attributes that will become clear in
the following commits.
This change is not backwards compatible, so I'm removing the open
current (really, v2) test. It'll be re-instated when we get to
https://github.com/mozilla/mentat/issues/194.
* Pre: Map TypedValue::Ref to TypedValue::Keyword in debug output.
* Part 3: Separate `schema_to_mutate` from the `schema` used to interpret.
This is just to keep track of the expected changes during
bootstrapping. I want bootstrap metadata mutations to flow through
the same code path as metadata mutations during regular transactions;
by differentiating the schema used for interpretation from the schema
that will be updated I expect to be able to apply bootstrap metadata
mutations to an empty schema and have things like materialized views
created (using the regular code paths).
This commit has been re-ordered for conceptual clarity, but it won't
compile because it references the metadata module. It's possible to
make it compile -- the functionality is there in the schema module --
but it's not worth the rebasing effort until after review (and
possibly not even then, since we'll squash down to a single commit to
land).
* Part 2: Maintain entids separately from idents.
In order to support historical idents, we need to distinguish the
"current" map from entid -> ident from the "complete historical" map
ident -> entid. This is what Datomic does; in Datomic, an ident is
never retracted (although it can be replaced). This approach is an
important part of allowing multiple consumers to share a schema
fragment as it migrates forward.
This fixes a limitation of the Clojure implementation, which did not
handle historical idents across knowledge base close and re-open.
The "entids" materialized view is naturally a slice of the "datoms"
table. The "idents" materialized view is a slice of the
"transactions" table. I hope that representing in this way, and
casting the problem in this light, might generalize to future
materialized views.
* Pre: Add DiffSet.
* Part 4: Collect mutations to a `Schema`.
I haven't taken your review comment about consuming AttributeBuilder
during each fluent function. If you read my response and still want
this, I'm happy to do it in review.
* Part 5: Handle :db/ident and :db.{install,alter}/attribute.
This "loops" the committed datoms out of the SQL store and back
through the metadata (schema, but in future also partition map)
processor. The metadata processor updates the schema and produces a
report of what changed; that report is then used to update the SQL
store. That update includes:
- the materialized views ("entids", "idents", and "schema");
- if needed, a subset of the datoms themselves (as flags change).
I've left a TODO for handling attribute retraction in the cases that
it makes sense. I expect that to be straight-forward.
* Review comment: Rename DiffSet to AddRetractAlterSet.
Also adds a little more commentary and a simple test.
* Review comment: Use ToIdent trait.
* Review comment: partially revert "Part 2: Maintain entids separately from idents."
This reverts commit 23a91df9c35e14398f2ddbd1ba25315821e67401.
Following our discussion, this removes the "entids" materialized
view. The next commit will remove historical idents from the "idents"
materialized view.
* Post: Use custom Either rather than std::result::Result.
This is not necessary, but it was suggested that we might be paying an
overhead creating Err instances while using error_chain. That seems
not to be the case, but this change shows that we don't actually use
any of the Result helper methods, so there's no reason to overload
Result. This change might avoid some future confusion, so I'm going
to land it anyway.
Signed-off-by: Nick Alexander <nalexander@mozilla.com>
* Review comment: Don't preserve historical idents.
* Review comment: More prepared statements when updating materialized views.
* Post: Test altering :db/cardinality and :db/unique.
These tests fail due to a Datomic limitation, namely that the marker
flag :db.alter/attribute can only be asserted once for an attribute!
That is, [:db.part/db :db.alter/attribute :attribute] will only be
transacted at most once. Since older versions of Datomic required the
:db.alter/attribute flag, I can only imagine they either never wrote
:db.alter/attribute to the store, or they handled it specially. I'll
need to remove the marker flag system from Mentat in order to address
this fundamental limitation.
* Post: Remove some more single quotes from error output.
* Post: Add assert_transact! macro to unwrap safely.
I was finding it very difficult to track unwrapping errors while
making changes, due to an underlying Mac OS X symbolication issue that
makes running tests with RUST_BACKTRACE=1 so slow that they all time
out.
* Post: Don't expect or recognize :db.{install,alter}/attribute.
I had this all working... except we will never see a repeated
`[:db.part/db :db.alter/attribute :attribute]` assertion in the store!
That means my approach would let you alter an attribute at most one
time. It's not worth hacking around this; it's better to just stop
expecting (and recognizing) the marker flags. (We have all the data
to distinguish the various cases that we need without the marker
flags.)
This brings Mentat in line with the thrust of newer Datomic versions,
but isn't compatible with Datomic, because (if I understand correctly)
Datomic automatically adds :db.{install,alter}/attribute assertions to
transactions.
I haven't purged the corresponding :db/ident and schema fragments just
yet:
- we might want them back
- we might want them in order to upgrade v1 and v2 databases to the
new on-disk layout we're fleshing out (v3?).
* Post: Don't make :db/unique :db.unique/* imply :db/index true.
This patch avoids a potential bug with the "schema" materialized view.
If :db/unique :db.unique/value implies :db/index true, then what
happens when you _retract_ :db.unique/value? I think Datomic defines
this in some way, but I really want the "schema" materialized view to
be a slice of "datoms" and not have these sort of ambiguities and
persistent effects. Therefore, to ensure that we don't retract a
schema characteristic and accidentally change more than we intended
to, this patch stops having any schema characteristic imply any other
schema characteristic(s). To achieve that, I added an
Option<Unique::{Value,Identity}> type to Attribute; this helps with
this patch, and also looks ahead to when we allow to retract
:db/unique attributes.
* Post: Allow to retract :db/ident.
* Post: Include more details about invalid schema changes.
The tests use strings, so they hide the chained errors which do in
fact provide more detail.
* Review comment: Fix outdated comment.
* Review comment: s/_SET/_SQL_LIST/.
* Review comment: Use a sub-select for checking cardinality.
This might be faster in practice.
* Review comment: Put `attribute::Unique` into its own namespace.
For queries like
```edn
[:find ?x :where [?x _ "hello"]]
[:find [?v ...] :where [_ ?a ?v]]
```
we'll query `all_datoms` to handle fulltext strings, which is expensive.
If `?a` is bound, we can avoid this — resolve any keyword binding,
ensure that the value is an attribute, and use the appropriate table.
This is necessary because we process patterns sequentially; a later
pattern might tell us the type of a variable (e.g., by having a
constant attribute), at which point we can do less work.
