Open CSV/Parquet/JSON files directly with DuckDB

sqlit/domains/connections/providers/duckdb/adapter.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+from pathlib import Path
 from typing import TYPE_CHECKING, Any
 
 from sqlit.domains.connections.providers.adapters.base import (
@@ -13,6 +14,11 @@
     TriggerInfo,
     resolve_file_path,
 )
+from sqlit.domains.connections.providers.duckdb.data_files import (
+    get_read_function,
+    sidecar_path_for,
+    table_name_for,
+)
 
 if TYPE_CHECKING:
     from sqlit.domains.connections.domain.config import ConnectionConfig
@@ -85,8 +91,45 @@ def connect(self, config: ConnectionConfig) -> Any:
         duckdb_any: Any = duckdb
         connect_args: dict[str, Any] = {}
         connect_args.update(config.extra_options)
+
+        read_fn = get_read_function(file_path)
+        if read_fn is not None:
+            return self._connect_data_file(
+                duckdb_any, file_path, read_fn, connect_args
+            )
+
         return duckdb_any.connect(str(file_path), **connect_args)
 
+    def _connect_data_file(
+        self,
+        duckdb_any: Any,
+        file_path: Path,
+        read_fn: str,
+        connect_args: dict[str, Any],
+    ) -> Any:
+        """Connect to a per-process sidecar `.duckdb` backed by a data file.
+
+        On first connect within a sqlit process the source file is loaded
+        into a real table so the user can run UPDATE/INSERT/DELETE against
+        it. Subsequent connects in the same process reuse the sidecar so
+        in-session edits persist across query Runs. The sidecar lives under
+        a PID-scoped temp dir; a new process gets a fresh load from source.
+        Writing back to the source is explicit: `COPY <table> TO '<path>'`.
+        """
+        sidecar = sidecar_path_for(file_path)
+        sidecar.parent.mkdir(parents=True, exist_ok=True)
+        needs_load = not sidecar.exists()
+
+        conn = duckdb_any.connect(str(sidecar), **connect_args)
+        if needs_load:
+            table = table_name_for(file_path)
+            path_literal = str(file_path).replace("'", "''")
+            conn.execute(
+                f'CREATE TABLE "{table}" AS '
+                f"SELECT * FROM {read_fn}('{path_literal}')"
+            )
+        return conn
+
     def get_databases(self, conn: Any) -> list[str]:
         """DuckDB doesn't support multiple databases - return empty list."""
         return []

sqlit/domains/connections/providers/duckdb/data_files.py

@@ -0,0 +1,106 @@
+"""Detection of data files queryable directly by DuckDB.
+
+DuckDB's `read_csv_auto`, `read_parquet`, `read_json_auto` etc. let you
+query a raw data file as if it were a table. When a sqlit DuckDB connection
+points at one of these files instead of a `.duckdb` database, the adapter:
+
+1. Picks a per-process sidecar `.duckdb` file in the OS temp dir.
+2. Loads the source file into a real TABLE in the sidecar on first connect.
+3. Lets the user CRUD the table freely; edits persist for the lifetime of
+   the sqlit process and are wiped on restart.
+4. Writing back to the source file is explicit (`COPY <table> TO '<path>'`).
+"""
+
+from __future__ import annotations
+
+import hashlib
+import os
+import re
+import tempfile
+from pathlib import Path
+
+
+# File extension -> DuckDB table function that can read it.
+_READ_FUNCTIONS: dict[str, str] = {
+    ".csv": "read_csv_auto",
+    ".tsv": "read_csv_auto",
+    ".parquet": "read_parquet",
+    ".pq": "read_parquet",
+    ".json": "read_json_auto",
+    ".jsonl": "read_json_auto",
+    ".ndjson": "read_json_auto",
+}
+
+# Allowed compression suffixes that wrap the data extensions above. DuckDB's
+# auto-readers transparently decompress these.
+_COMPRESSION_SUFFIXES: frozenset[str] = frozenset({".gz", ".zst", ".bz2"})
+
+
+def get_read_function(path: Path) -> str | None:
+    """Return the DuckDB table function for this file, or None if not a known
+    data file extension.
+
+    Handles compressed forms like `.csv.gz` by looking past the compression
+    suffix.
+    """
+    suffixes = [s.lower() for s in path.suffixes]
+    if not suffixes:
+        return None
+
+    last = suffixes[-1]
+    if last in _COMPRESSION_SUFFIXES:
+        if len(suffixes) >= 2:
+            return _READ_FUNCTIONS.get(suffixes[-2])
+        return None
+    return _READ_FUNCTIONS.get(last)
+
+
+def is_data_file(path: Path) -> bool:
+    """True if the file extension is one DuckDB can query directly."""
+    return get_read_function(path) is not None
+
+
+def table_name_for(path: Path) -> str:
+    """Build a SQL-safe table name from a file path basename.
+
+    Strips the data and (optional) compression extension, then sanitizes
+    non-identifier characters to underscores.
+
+    Examples:
+        sales.csv         -> sales
+        sales-2024.csv    -> sales_2024
+        events.json.gz    -> events
+        123-data.parquet  -> _123_data
+    """
+    stem = path.name
+    # Strip compression suffix if present.
+    lower = stem.lower()
+    for comp in _COMPRESSION_SUFFIXES:
+        if lower.endswith(comp):
+            stem = stem[: -len(comp)]
+            break
+    # Strip data extension.
+    dot = stem.rfind(".")
+    if dot > 0:
+        stem = stem[:dot]
+
+    sanitized = re.sub(r"[^A-Za-z0-9_]+", "_", stem).strip("_")
+    if not sanitized:
+        sanitized = "data"
+    if sanitized[0].isdigit():
+        sanitized = "_" + sanitized
+    return sanitized.lower()
+
+
+def sidecar_path_for(source_path: Path) -> Path:
+    """Per-process scratch `.duckdb` path for a data-file source.
+
+    Each sqlit process gets its own directory under the OS temp dir. The
+    sidecar persists for the lifetime of the process so edits within a
+    sqlit session survive across query Runs. A fresh process gets a fresh
+    sidecar, so source-file changes are picked up on restart and unsaved
+    edits are wiped (the user opted into "re-load from source each time").
+    """
+    digest = hashlib.sha1(str(source_path.resolve()).encode()).hexdigest()[:16]
+    base = Path(tempfile.gettempdir()) / f"sqlit-{os.getpid()}" / "data-files"
+    return base / f"{digest}.duckdb"