This PR implements the `Sequence` and `SequenceTest` opcodes, although
does not yet add plumbing to emit the latter.
SQLite has two distinct mechanisms that determine the final row order
with aggregates:
Traversal order of GROUP BY, and ORDER BY tiebreaking. When ORDER BY
contains only aggregate expressions and/or constants, SQLite has no
extra tiebreak key, but when ORDER BY mixes aggregate and non-aggregate
terms, SQLite adds an implicit, stable row `sequence` so “ties” respect
the input order.
This PR also fixes an issue with a query like the following:
```sql
SELECT u.first_name, COUNT(*) AS c
FROM users u
JOIN orders o ON o.user_id = u.id
GROUP BY u.first_name
ORDER BY c DESC;
```
Because ORDER BY has only an aggregate (COUNT(*) DESC) and no non-
aggregate terms, SQLite traverses the group key (u.first_name) in DESC
order in this case, so ties on c naturally appear with group keys in
descending order.
Previously tursodb would return the group key sorted in ASC order,
because it was used in all cases as the default
Closes#3287
Adds initial support for window functions. For now, only existing
aggregate functions can be used as window functions—no specialized
window-specific functions are supported yet.
Currently, only the default frame definition is implemented:
RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW EXCLUDE NO OTHERS.
The MakeRecord instruction now accepts an optional affinity_str
parameter that applies column-specific type conversions before creating
records. When provided, the affinity string is applied
character-by-character to each register using the existing
apply_affinity_char() function, matching SQLite's behavior.
Fixes#2040Fixes#2041
This fairly long commit implements persistence for materialized view.
It is hard to split because of all the interdependencies between components,
so it is a one big thing. This commit message will at least try to go into
details about the basic architecture.
Materialized Views as tables
============================
Materialized views are now a normal table - whereas before they were a virtual
table. By making a materialized view a table, we can reuse all the
infrastructure for dealing with tables (cursors, etc).
One of the advantages of doing this is that we can create indexes on view
columns. Later, we should also be able to write those views to separate files
with ATTACH write.
Materialized Views as Zsets
===========================
The contents of the table are a ZSet: rowid, values, weight. Readers will
notice that because of this, the usage of the ZSet data structure dwindles
throughout the codebase. The main difference between our materialized ZSet and
the standard DBSP ZSet, is that obviously ours is backed by a BTree, not a Hash
(since SQLite tables are BTrees)
Aggregator State
================
In DBSP, the aggregator nodes also have state. To store that state, there is a
second table. The table holds all aggregators in the view, and there is one
table per view. That is __turso_internal_dbsp_state_{view_name}. The format of
that table is similar to a ZSet: rowid, serialized_values, weight. We serialize
the values because there will be many aggregators in the table. We can't rely
on a particular format for the values.
The Materialized View Cursor
============================
Reading from a Materialized View essentially means reading from the persisted
ZSet, and enhancing that with data that exists within the transaction.
Transaction data is ephemeral, so we do not materialize this anywhere: we have
a carefully crafted implementation of seek that takes care of merging weights
and stitching the two sets together.
SQLite does not store the rowid alias column in the record at all
when it is a rowid alias, because the rowid is always stored anyway
in the record header.
In case an ORDER BY column exactly matches a result column in the SELECT,
the insertion of the result column into the ORDER BY sorter can be skipped
because it's already necessarily inserted as a sorting column.
For this reason we have a mapping to know what index a given result column
has in the order by sorter.
This commit makes that mapping much simpler.
This change connects virtual tables with the query optimizer.
The optimizer now considers virtual tables during join order search
and invokes their best_index callbacks to determine feasible access
paths.
Currently, this is not a visible change, since none of the existing
extensions return information indicating that a plan is invalid.
Different scan parameters are required for different table types.
Currently, index and iteration direction are only used by B-tree tables,
while the remaining table types don’t require any parameters. Planning
access to virtual tables, however, will require passing additional
information from the planner, such as the virtual table index (distinct
from a B-tree index) and the constraints that must be forwarded to the
`filter` method.
The `best_index` implementation now returns a ResultCode along with the
IndexInfo. This allows it to signal specific outcomes, such as errors or
constraint violations. This change aligns better with SQLite’s xBestIndex
contract, where cases like missing constraints or invalid combinations of
constraints must not result in a valid plan.
Previously, there were two ways to indicate that a constraint should not
be passed to the filter function: setting `argv_index` to `None` or to
a value less than 1. This was redundant, so now only `None` is used.
Additional changes:
- Update IndexInfo documentation to clarify that constraint_usages must
have exact 1:1 correspondence with input ConstraintInfo array. The code
translating constraints into VFilter arguments heavily relies on this.
- Fix best_index implementation in test extension to comply with new
validation requirements by returning usage entry for each constraint
Support for attaching databases. The main difference from SQLite is that
we support an arbitrary number of attached databases, and we are not
bound to just 100ish.
We for now only support read-only databases. We open them as read-only,
but also, to keep things simple, we don't patch any of the insert
machinery to resolve foreign tables. So if an insert is tried on an
attached database, it will just fail with a "no such table" error - this
is perfect for now.
The code in core/translate/attach.rs is written by Claude, who also
played a key part in the boilerplate for stuff like the .databases
command and extending the pragma database_list, and also aided me in
the test cases.
There's no such thing as a read-only connection.
In a normal connection, you can have many attached databases. Some
r/o, some r/w.
To properly fix that, we also need to fix the OpenWrite opcode. Right
now we are passing a name, which is the name of the table. That
parameter is not used anywhere. That is also not what the SQLite opcode
specifies. Same as OpenRead, the p3 register should be the database
index.
With that change, we can - for now - pass the index 0, which is all
we support anyway, and then use that to test if we are r/o.
Two of the opcodes we implement (OpenRead and Transaction) should have
an opcode specifying the database to use, but they don't.
Add it, and for now always use 0 (the main database).
With this change, the following two queries are considered equivalent:
```sql
SELECT value FROM generate_series(5, 50);
SELECT value FROM generate_series WHERE start = 5 AND stop = 50;
```
Arguments passed in parentheses to the virtual table name are now
matched to hidden columns.
Column references are still not supported as table-valued function
arguments. The only difference is that previously, a query like:
```sql
SELECT one.value, series.value
FROM (SELECT 1 AS value) one, generate_series(one.value, 3) series;
```
would cause a panic. Now, it returns a proper error message instead.
Adding support for column references is more nuanced for two main
reasons:
- We need to ensure that in joins where a TVF depends on other tables,
those other tables are processed first. For example, in:
```sql
SELECT one.value, series.value
FROM generate_series(one.value, 3) series, (SELECT 1 AS value) one;
```
the one table must be processed by the top-level loop, and series must
be nested.
- For outer joins involving TVFs, the arguments must be treated as ON
predicates, not WHERE predicates.
We need to enumerate first and filter afterward — not the other way
around — because we later use the indexes produced by `enumerate` to
access the original `predicates` slice.
Fixes DELETE not emitting conditional jumps at all if the associated
WhereTerm is a constant, e.g.
```sql
limbo> create table t(x);
limbo> explain DELETE FROM t WHERE 5-5;
addr opcode p1 p2 p3 p4 p5 comment
---- ----------------- ---- ---- ---- ------------- -- -------
0 Init 0 7 0 0 Start at 7
1 OpenWrite 0 2 0 0 root=2; t
2 Rewind 0 6 0 0 Rewind table t
3 RowId 0 1 0 0 r[1]=t.rowid
4 Delete 0 0 0 0
5 Next 0 3 0 0
6 Halt 0 0 0 0
7 Transaction 0 1 0 0 write=true
8 Goto 0 1 0 0
```
I was adding more stuff to the simulator in a Branch of mine, and I
caught this error with delete. Upstreaming the fix here. As we do with
Update, I added the translation step for the `WhereTerms` of the query.
Edit: Closes#1732. Closes#1733. Closes#1734. Closes#1735. Closes
#1736. Closes#1738. Closes#1739. Closes#1740.
Edit: Also pushes constant where term translation to `init_loop` for
Update and Select as well.
Reviewed-by: Jussi Saurio <jussi.saurio@gmail.com>
Closes#1746
Previously, the logic for collecting non-aggregate columns was duplicated
across multiple locations and implemented inconsistently. This caused a
bug that was revealed by the refactoring in this commit (see the added
test).
