Fix Power BI DirectQuery slicers on text, interval, time and timestamptz columns

.github/workflows/build.yaml

@@ -310,8 +310,8 @@ jobs:
       # per layer: once a single layer exhausts it, the pull fails and takes
       # every other image with it. Layers that did arrive stay in the local
       # content store, so a second attempt refetches only what is missing and
-      # costs seconds. The spooling leg is the one that reaches quay.io for
-      # MinIO, and that is where the dropped connections have been seen.
+      # costs seconds. The spooling leg pulls the most, MinIO on top of the
+      # core images, so it is the one most exposed to a dropped connection.
       - name: Pull the stack images
         # From the stack directory, as scripts/lib.sh's `compose` wrapper does.
         # Compose derives the project name from the compose file's directory,

CHANGELOG.md

@@ -7,6 +7,46 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Changed
+
+- `timestamp with time zone` values are delivered as wall time in the session
+  time zone instead of in UTC: `TimeZone=` when set, otherwise the
+  coordinator's default, and whatever a later `SET TIME ZONE` chose. Power BI
+  folds the value it showed back as a plain `TIMESTAMP` literal, which Trino
+  reads in the session zone, so a DirectQuery slicer on such a column only
+  selected its rows in a UTC session. Applications reading these columns see
+  different values unless the session is UTC. `time with time zone` values
+  follow the same zone, at its current offset because a time has no date, which
+  is how Trino itself casts `TIME` to `TIME WITH TIME ZONE`.
+- The Power BI connector reports that Trino cannot convert `TIME` to
+  `TIMESTAMP` (`SQL_CONVERT_TIME` without `SQL_CVT_TIMESTAMP`). A DirectQuery
+  slicer on a time column folded to a comparison against Power BI's base date,
+  30 December 1899, while Trino anchors a cast time on the current date, so the
+  report silently showed no rows. Power BI now refuses that fold with a visible
+  error instead. The README describes a view-based workaround.
+
+### Fixed
+
+- `SQLGetTypeInfo` lists the plain type first among the rows that share a
+  `DATA_TYPE`: `VARCHAR` for `SQL_WVARCHAR`, `TIME` for `SQL_TYPE_TIME` and
+  `TIMESTAMP` for `SQL_TYPE_TIMESTAMP`, as the spec's "how closely the data type
+  maps" ordering requires. The rows were sorted by name, so `SQL_WVARCHAR` led
+  with `INTERVAL DAY TO SECOND`, and Power Query, which takes the first row as
+  its `CAST` target, folded a DirectQuery slicer on a text column into
+  `CAST(... AS INTERVAL DAY TO SECOND)`, which Trino rejected.
+- The Power BI connector quotes text constants and doubles any `'` in them.
+  Power Query hands the constant over unquoted, so a slicer value folded into
+  `CAST(hello world as VARCHAR)` and failed. Together with the `SQLGetTypeInfo`
+  fix above this makes DirectQuery slicers on text columns work, so the driver
+  and the connector have to be upgraded together.
+- `INTERVAL YEAR TO MONTH` and `INTERVAL DAY TO SECOND` columns read as text now
+  return Trino's own rendering, the same text `CAST(... AS VARCHAR)` produces
+  (`-1-0`, `0 00:00:00.500`). They were parsed into fields and re-rendered
+  (`-1-00`, `0 00:00:00.5`), so a Power BI DirectQuery slicer on an interval
+  column, which folds to `cast(col as VARCHAR) = '<shown value>'`, silently
+  selected no rows. Reading these columns as `SQL_C_INTERVAL_*` still works:
+  stackable-odbc-core now converts interval text to those C types.
+
 ## [0.1.2] — 2026-09-01
 
 ### Changed

Cargo.toml

@@ -38,7 +38,7 @@ snafu = "0.9"
 # reviewable edit to this line instead of whatever the branch happens to point
 # at. To build against a local checkout, use a `[patch]` in your own
 # `.cargo/config.toml` rather than editing this line; see CONTRIBUTING.md.
-stackable-odbc-core = { git = "https://github.com/stackabletech/stackable-odbc-core.git", tag = "v0.1.0" }
+stackable-odbc-core = { git = "https://github.com/stackabletech/stackable-odbc-core.git", tag = "v0.1.1" }
 # `time` is needed directly by `query_all_rows_within`, which bounds the login
 # round trip with `tokio::time::timeout`. It resolves without being declared,
 # because reqwest enables it, but a direct use must not rely on another crate's
@@ -60,7 +60,7 @@ serial_test = "4"
 # offline FFI tests need. It is default-off because it is test code, so it is
 # enabled here rather than on the [dependencies] entry above -- that keeps it
 # out of the shipped cdylib.
-stackable-odbc-core = { git = "https://github.com/stackabletech/stackable-odbc-core.git", tag = "v0.1.0", features = ["test-support"] }
+stackable-odbc-core = { git = "https://github.com/stackabletech/stackable-odbc-core.git", tag = "v0.1.1", features = ["test-support"] }
 
 [lints.clippy]
 unwrap_in_result = "deny"

README.md

@@ -78,6 +78,11 @@ proper entry in the **Get Data** dialog instead of the generic ODBC one.
 2. In **File > Options > Security**, allow any extension to load.
 3. Restart Power BI Desktop. **Stackable Trino** now appears under **Get Data**.
 
+For incremental refresh on a `timestamp with time zone` column, Trino compares
+the `RangeStart` and `RangeEnd` bounds in the session time zone (`TimeZone`),
+so partition boundaries follow that zone. Changing `TimeZone` on a dataset that
+already has partitions moves the boundaries.
+
 ### Your first query
 
 Assuming a Trino instance is reachable on the given host and port:
@@ -148,7 +153,7 @@ The authoritative list is `src/backend/types/connect_params.rs`.
 | `Roles` | No | Authorisation role per catalog, `{catalog:role;catalog2:ALL}` |
 | `SessionUser` | No | User statements run as, while `User` still authenticates. JDBC's `sessionUser` |
 | `Path` | No | Default SQL path for resolving unqualified function names |
-| `TimeZone` | No | IANA session time zone (`Europe/Berlin`). Unset leaves the coordinator's |
+| `TimeZone` | No | IANA session time zone (`Europe/Berlin`). Unset leaves the coordinator's default. `timestamp with time zone` and `time with time zone` values are delivered as wall time in the session zone, which a later `SET TIME ZONE` changes |
 | `Locale` | No | Locale for locale-dependent formatting, sent as `X-Trino-Language` |
 | `ClientInfo` | No | Free-form client metadata Trino records against the query |
 | `TraceToken` | No | Correlation token Trino records against the query |
@@ -259,6 +264,19 @@ ignored, so the tool can react instead of trusting a wrong answer.
   somewhere you never asked for. Set `Catalog` when you connect.
 - **Row and field size limits are not faked.** Trino can only limit a result set
   through `LIMIT` in the SQL you wrote.
+- **Power BI cannot filter on a time column in DirectQuery.** Picking a value
+  in a slicer on a `time` column fails with "We couldn't fold the expression to
+  the data source". Power BI would filter by casting the column to a timestamp
+  and comparing it with that time on 30 December 1899, while Trino, like ODBC
+  itself, puts a cast time on today's date, so the filter could never match.
+  The connector declares the cast unsupported so you see an error instead of an
+  empty report. To slice on a time of day, expose it as text in a Trino view,
+  for example `CAST(col_time AS VARCHAR) AS col_time_text`, and slice on that
+  column. Selecting "(Blank)" still works. Users who cannot create views can
+  switch the table to Import mode, where the slicer filters Power BI's own copy
+  of the data. That copy is only as fresh as its last refresh, and Power BI
+  treats a blank time as equal to midnight, so selecting 12:00:00 AM also shows
+  rows without a time.
 - **One isolation level.** Trino catalogs disagree about which levels they
   accept, so the driver offers the one they all support and refuses the rest up
   front, rather than letting a query fail later for a reason nobody can see.
@@ -307,6 +325,22 @@ the coordinator's chain is refused even when the machine trusts that chain.
 **Only the first session property applies.** Wrap the value in braces. See
 [Values that contain a semicolon](#values-that-contain-a-semicolon).
 
+**A Power BI slicer misses a timestamp from the night the clocks go back.**
+`timestamp with time zone` values are shown in the session time zone, and in
+the hour that repeats, two instants share one wall time. Power BI filters on the
+wall time it showed and Trino reads that as the later of the two instants, so a
+value from the first of the repeated hours is not selected.
+
+**Large reads fail with `ABANDONED_QUERY` or `Query not found`.** Trino
+abandons a query whose results the client has not fetched within
+`query.client.timeout` (5 minutes by default), and later forgets it altogether,
+after which the next fetch returns `404 Not Found: Query not found`. A client
+that reads slowly enough gets there on a large result. One cause on
+Windows is ODBC tracing left switched on: it writes every call to a file and
+slows reads down considerably. Turn it off in the ODBC Data Source
+Administrator (Tracing tab, **Stop Tracing Now**), including on an on-premises
+data gateway.
+
 **The browser login never opens.** Some tools, `pyodbc` among them, tell the
 driver it may not display anything. The driver reports this rather than hanging.
 Use `AccessToken` with those tools, or connect through one that allows a prompt.

connector/StackableTrinoODBC.pq

@@ -307,10 +307,10 @@ StackableTrinoODBCImpl = (
             SupportsTop = false
         ],
 
-        // Nothing is overridden. An override here silently wins over
-        // SQLGetInfoW and cannot be corrected by fixing the driver, so the
-        // record is reserved for what the driver gets wrong, and this group it
-        // answers honestly: SQL_SQL92_PREDICATES, SQL_AGGREGATE_FUNCTIONS,
+        // One entry is overridden, SQL_CONVERT_TIME, below. An override here
+        // silently wins over SQLGetInfoW and cannot be corrected by fixing the
+        // driver, so the record is otherwise reserved for what the driver gets
+        // wrong, and this group it answers honestly: SQL_SQL92_PREDICATES, SQL_AGGREGATE_FUNCTIONS,
         // SQL_SQL92_RELATIONAL_JOIN_OPERATORS, SQL_SQL92_VALUE_EXPRESSIONS and
         // SQL_IDENTIFIER_QUOTE_CHAR.
         //
@@ -327,7 +327,22 @@ StackableTrinoODBCImpl = (
         // Nothing Power Query generates is lost: comparison, IN,
         // LIKE, BETWEEN, IS NULL, EXISTS and the four join types are all in
         // the driver's ungated set.
-        SQLGetInfo = defaultConfig[SQLGetInfo],
+        //
+        // SQL_CONVERT_TIME is the driver's answer (every SQL_CVT_* bit) minus
+        // SQL_CVT_TIMESTAMP, and that is a deliberate misreport: Trino can
+        // cast TIME to TIMESTAMP. It anchors the time on the current date, as
+        // ODBC's own conversion tables do ("SQL to C: Time", footnote [c]),
+        // while Power BI folds a time slicer to
+        // `cast(col as TIMESTAMP) = CAST('1899-12-30 hh:mm:ss' as TIMESTAMP)`,
+        // anchoring on its own base date. That filter ran and silently matched
+        // no row. Without the bit, Power BI refuses the fold and says so ("We
+        // couldn't fold the expression to the data source"), which is the
+        // honest outcome for a filter Trino cannot evaluate the way Power BI
+        // means it. test_folding_contract.py pins the value to the driver's
+        // answer and checks Trino's anchor is still the current date.
+        SQLGetInfo = defaultConfig[SQLGetInfo] & [
+            SQL_CONVERT_TIME = 0x01FDFFFF
+        ],
 
         SQLColumns = (catalogName, schemaName, tableName, columnName, source) =>
             source,
@@ -394,12 +409,16 @@ StackableTrinoODBCImpl = (
                     //             rendering it as Trino's `X'..'` literal
                     //             cannot be verified without Power BI Desktop.
                     //
+                    // Power Query hands a text-valued constant to the visitor
+                    // bare, not quoted: confirmed in Power BI Desktop on
+                    // 2026-10-07, where `Cast(_, "VARCHAR")` folded a slicer
+                    // into `CAST(hello world as VARCHAR)`. A text entry therefore
+                    // quotes the value itself and doubles any single quote in
+                    // it, or `O'Brien` ends the literal early.
+                    //
                     // TODO: add UUID, JSON, TIME WITH TIME ZONE and TIMESTAMP
-                    // WITH TIME ZONE. Each has a valid Trino cast, but the
-                    // rendering turns on whether Power Query hands a
-                    // text-valued constant to the visitor already quoted, which
-                    // the VARCHAR entry below also rests on. Confirm that in
-                    // Power BI Desktop, then add all four.
+                    // WITH TIME ZONE, now that the quoting question above is
+                    // settled. Each has a valid Trino cast.
                     Visitor = [
                         DECIMAL = each Cast(_, "DECIMAL"),
                         INTEGER = each Cast(_, "INTEGER"),
@@ -410,7 +429,7 @@ StackableTrinoODBCImpl = (
                         DOUBLE = each Cast(_, "DOUBLE PRECISION"),
                         BOOLEAN = each Cast(_, "BOOLEAN"),
                         DATE = each Cast(Quote(Date.ToText(_, "yyyy-MM-dd")), "DATE"),
-                        VARCHAR = each Cast(_, "VARCHAR"),
+                        VARCHAR = each Cast(Quote(Text.Replace(_, "'", "''")), "VARCHAR"),
                         // `fffffff`, not `sssssss`: in a custom format string
                         // `s` is the second and `f` is the fractional second,
                         // so the latter spelling renders the second eight times

fuzz/README.md

@@ -20,8 +20,8 @@ one: a value a coordinator legitimately sent must fail safe, and where a
 release build has no overflow checks the same defect returns a wrong answer
 instead of an error.
 
-- `json_value` covers `json_to_column_value` and the dozen temporal, interval
-  and decimal scanners under it. This is the half of the read path core does
+- `json_value` covers `json_to_column_value` and the temporal and decimal
+  scanners under it. This is the half of the read path core does
   not see: core fuzzes `write_column_value`, which turns the resulting
   `ColumnValue` into the caller's buffer, and nothing covered the step that
   produces it.

fuzz/fuzz_targets/json_value.rs

@@ -11,8 +11,8 @@ use trino_rust_client::{TrinoFloat, TrinoInt, TrinoTy};
 //
 // stackable-odbc-core already fuzzes the second half (`write_column_value`,
 // ColumnValue -> the caller's buffer). Nothing covered the step before it,
-// which is where this crate's temporal, interval and decimal parsers live:
-// roughly a dozen hand-written scanners over text a Trino coordinator chose.
+// which is where this crate's temporal and decimal parsers live: hand-written
+// scanners over text a Trino coordinator chose.
 // Every one of them runs on the server's side of the trust boundary.
 //
 // The property is that no input panics. A panic here is caught at the FFI

integration-tests/scripts/gen-trino-config.sh

@@ -19,7 +19,12 @@ set -euo pipefail
 source "$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)/lib.sh"
 
 OUT="$GENERATED/trino"
-rm -rf "$OUT"
+# Emptied in place, never removed: a running coordinator bind-mounts this
+# directory, and replacing it would leave the container holding the deleted
+# one, empty, so a restart would find no config at all. setup.sh recreates
+# Trino when the assembled content changes (TRINO_CONFIG_HASH).
+mkdir -p "$OUT"
+find "$OUT" -mindepth 1 -delete
 mkdir -p "$OUT/catalog"
 
 cp -r "$STACK_DIR/trino/base/." "$OUT/"