mcp-stata 1.18.0__cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl

mcp_stata/ui_http.py

@@ -0,0 +1,998 @@
+import hashlib
+import io
+import json
+import secrets
+import threading
+import time
+import uuid
+import logging
+from dataclasses import dataclass
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from typing import Any, Callable, Optional
+
+from .stata_client import StataClient
+from .config import (
+    DEFAULT_HOST,
+    DEFAULT_PORT,
+    MAX_ARROW_LIMIT,
+    MAX_CHARS,
+    MAX_LIMIT,
+    MAX_REQUEST_BYTES,
+    MAX_VARS,
+    TOKEN_TTL_S,
+    VIEW_TTL_S,
+)
+
+
+logger = logging.getLogger("mcp_stata")
+
+try:
+    from .native_ops import argsort_numeric as _native_argsort_numeric
+    from .native_ops import argsort_mixed as _native_argsort_mixed
+except Exception:
+    _native_argsort_numeric = None
+    _native_argsort_mixed = None
+
+
+def _try_native_argsort(
+    table: Any,
+    sort_cols: list[str],
+    descending: list[bool],
+    nulls_last: list[bool],
+) -> list[int] | None:
+    if _native_argsort_numeric is None and _native_argsort_mixed is None:
+        return None
+    try:
+        import pyarrow as pa
+        import numpy as np
+
+        is_string: list[bool] = []
+        cols: list[object] = []
+        for col in sort_cols:
+            arr = table.column(col).combine_chunks()
+            if pa.types.is_string(arr.type) or pa.types.is_large_string(arr.type):
+                is_string.append(True)
+                cols.append(arr.to_pylist())
+                continue
+            if not (pa.types.is_floating(arr.type) or pa.types.is_integer(arr.type)):
+                return None
+            is_string.append(False)
+            np_arr = arr.to_numpy(zero_copy_only=False)
+            if np_arr.dtype != np.float64:
+                np_arr = np_arr.astype(np.float64, copy=False)
+            # Normalize Stata missing values for numeric columns
+            np_arr = np.where(np_arr > 8.0e307, np.nan, np_arr)
+            cols.append(np_arr)
+
+        if "_n" not in table.column_names:
+            return None
+        obs = table.column("_n").to_numpy(zero_copy_only=False).astype(np.int64, copy=False)
+
+        if all(not flag for flag in is_string) and _native_argsort_numeric is not None:
+            idx = _native_argsort_numeric(cols, descending, nulls_last)
+        elif _native_argsort_mixed is not None:
+            idx = _native_argsort_mixed(cols, is_string, descending, nulls_last)
+        else:
+            return None
+
+        return [int(x) for x in (obs[idx] - 1).tolist()]
+    except Exception:
+        return None
+
+
+def _get_sorted_indices_polars(
+    table: Any,
+    sort_cols: list[str],
+    descending: list[bool],
+    nulls_last: list[bool],
+) -> list[int]:
+    import polars as pl
+
+    df = pl.from_arrow(table)
+    # Normalize Stata missing values for numeric columns
+    exprs = []
+    for col, dtype in zip(df.columns, df.dtypes):
+        if col == "_n":
+            exprs.append(pl.col(col))
+            continue
+        if dtype in (pl.Float32, pl.Float64):
+            exprs.append(
+                pl.when(pl.col(col) > 8.0e307)
+                .then(None)
+                .otherwise(pl.col(col))
+                .alias(col)
+            )
+        else:
+            exprs.append(pl.col(col))
+    df = df.select(exprs)
+
+    try:
+        # Use expressions for arithmetic to avoid eager Series-scalar conversion issues
+        # that have been observed in some environments with Int64 dtypes.
+        res = df.select(
+            idx=pl.arg_sort_by(
+                [pl.col(c) for c in sort_cols],
+                descending=descending,
+                nulls_last=nulls_last,
+            ),
+            zero_based_n=pl.col("_n") - 1
+        )
+        return res["zero_based_n"].take(res["idx"]).to_list()
+    except Exception:
+        # Fallback to eager sort if arg_sort_by fails or has issues
+        return (
+            df.sort(by=sort_cols, descending=descending, nulls_last=nulls_last)
+            .select(pl.col("_n") - 1)
+            .to_series()
+            .to_list()
+        )
+
+
+
+
+def _stable_hash(payload: dict[str, Any]) -> str:
+    return hashlib.sha1(json.dumps(payload, sort_keys=True).encode("utf-8")).hexdigest()
+
+
+@dataclass
+class UIChannelInfo:
+    base_url: str
+    token: str
+    expires_at: int
+
+
+@dataclass
+class ViewHandle:
+    view_id: str
+    dataset_id: str
+    frame: str
+    filter_expr: str
+    obs_indices: list[int]
+    filtered_n: int
+    created_at: float
+    last_access: float
+
+
+class UIChannelManager:
+    def __init__(
+        self,
+        client: StataClient,
+        *,
+        host: str = DEFAULT_HOST,
+        port: int = DEFAULT_PORT,
+        token_ttl_s: int = TOKEN_TTL_S,
+        view_ttl_s: int = VIEW_TTL_S,
+        max_limit: int = MAX_LIMIT,
+        max_vars: int = MAX_VARS,
+        max_chars: int = MAX_CHARS,
+        max_request_bytes: int = MAX_REQUEST_BYTES,
+        max_arrow_limit: int = MAX_ARROW_LIMIT,
+    ):
+        self._client = client
+        self._host = host
+        self._port = port
+        self._token_ttl_s = token_ttl_s
+        self._view_ttl_s = view_ttl_s
+        self._max_limit = max_limit
+        self._max_vars = max_vars
+        self._max_chars = max_chars
+        self._max_request_bytes = max_request_bytes
+        self._max_arrow_limit = max_arrow_limit
+
+        self._lock = threading.Lock()
+        self._httpd: ThreadingHTTPServer | None = None
+        self._thread: threading.Thread | None = None
+
+        self._token: str | None = None
+        self._expires_at: int = 0
+
+        self._dataset_version: int = 0
+        self._dataset_id_cache: str | None = None
+        self._dataset_id_cache_at_version: int = -1
+
+        self._views: dict[str, ViewHandle] = {}
+        self._last_sort_spec: tuple[str, ...] | None = None
+        self._last_sort_dataset_id: str | None = None
+        self._sort_index_cache: dict[tuple[str, tuple[str, ...]], list[int]] = {}
+        self._sort_cache_order: list[tuple[str, tuple[str, ...]]] = []
+        self._sort_cache_max_entries: int = 4
+        self._sort_table_cache: dict[tuple[str, tuple[str, ...]], Any] = {}
+        self._sort_table_order: list[tuple[str, tuple[str, ...]]] = []
+        self._sort_table_max_entries: int = 2
+
+    def notify_potential_dataset_change(self) -> None:
+        with self._lock:
+            self._dataset_version += 1
+            self._dataset_id_cache = None
+            self._views.clear()
+            self._last_sort_spec = None
+            self._last_sort_dataset_id = None
+            self._sort_index_cache.clear()
+            self._sort_cache_order.clear()
+            self._sort_table_cache.clear()
+            self._sort_table_order.clear()
+
+    @staticmethod
+    def _normalize_sort_spec(sort_spec: list[str]) -> tuple[str, ...]:
+        normalized: list[str] = []
+        for spec in sort_spec:
+            if not isinstance(spec, str) or not spec:
+                raise ValueError(f"Invalid sort specification: {spec!r}")
+            raw = spec.strip()
+            if not raw:
+                raise ValueError(f"Invalid sort specification: {spec!r}")
+            sign = "-" if raw.startswith("-") else "+"
+            varname = raw.lstrip("+-")
+            if not varname:
+                raise ValueError(f"Invalid sort specification: {spec!r}")
+            normalized.append(f"{sign}{varname}")
+        return tuple(normalized)
+
+    def _get_cached_sort_indices(
+        self, dataset_id: str, sort_spec: tuple[str, ...]
+    ) -> list[int] | None:
+        key = (dataset_id, sort_spec)
+        with self._lock:
+            cached = self._sort_index_cache.get(key)
+            if cached is None:
+                return None
+            # refresh LRU order
+            if key in self._sort_cache_order:
+                self._sort_cache_order.remove(key)
+            self._sort_cache_order.append(key)
+            return cached
+
+    def _set_cached_sort_indices(
+        self, dataset_id: str, sort_spec: tuple[str, ...], indices: list[int]
+    ) -> None:
+        key = (dataset_id, sort_spec)
+        with self._lock:
+            if key in self._sort_index_cache:
+                self._sort_cache_order.remove(key)
+            self._sort_index_cache[key] = indices
+            self._sort_cache_order.append(key)
+            while len(self._sort_cache_order) > self._sort_cache_max_entries:
+                evict = self._sort_cache_order.pop(0)
+                self._sort_index_cache.pop(evict, None)
+
+    def _get_cached_sort_table(
+        self, dataset_id: str, sort_cols: tuple[str, ...]
+    ) -> Any | None:
+        key = (dataset_id, sort_cols)
+        with self._lock:
+            cached = self._sort_table_cache.get(key)
+            if cached is None:
+                return None
+            if key in self._sort_table_order:
+                self._sort_table_order.remove(key)
+            self._sort_table_order.append(key)
+            return cached
+
+    def _set_cached_sort_table(
+        self, dataset_id: str, sort_cols: tuple[str, ...], table: Any
+    ) -> None:
+        key = (dataset_id, sort_cols)
+        with self._lock:
+            if key in self._sort_table_cache:
+                self._sort_table_order.remove(key)
+            self._sort_table_cache[key] = table
+            self._sort_table_order.append(key)
+            while len(self._sort_table_order) > self._sort_table_max_entries:
+                evict = self._sort_table_order.pop(0)
+                self._sort_table_cache.pop(evict, None)
+
+    def _get_sort_table(self, dataset_id: str, sort_cols: list[str]) -> Any:
+        sort_cols_key = tuple(sort_cols)
+        cached = self._get_cached_sort_table(dataset_id, sort_cols_key)
+        if cached is not None:
+            return cached
+
+        state = self._client.get_dataset_state()
+        n = int(state.get("n", 0) or 0)
+        if n <= 0:
+            return None
+
+        # Pull full columns once via Arrow stream (Stata -> Arrow), then sort in Polars.
+        arrow_bytes = self._client.get_arrow_stream(
+            offset=0,
+            limit=n,
+            vars=sort_cols,
+            include_obs_no=True,
+            obs_indices=None,
+        )
+
+        import pyarrow as pa
+
+        with pa.ipc.open_stream(io.BytesIO(arrow_bytes)) as reader:
+            table = reader.read_all()
+
+        self._set_cached_sort_table(dataset_id, sort_cols_key, table)
+        return table
+
+    def get_channel(self) -> UIChannelInfo:
+        self._ensure_http_server()
+        with self._lock:
+            self._ensure_token()
+            assert self._httpd is not None
+            port = self._httpd.server_address[1]
+            base_url = f"http://{self._host}:{port}"
+            return UIChannelInfo(base_url=base_url, token=self._token or "", expires_at=self._expires_at)
+
+    def capabilities(self) -> dict[str, bool]:
+        return {"dataBrowser": True, "filtering": True, "sorting": True, "arrowStream": True}
+
+    def current_dataset_id(self) -> str:
+        with self._lock:
+            if self._dataset_id_cache is not None and self._dataset_id_cache_at_version == self._dataset_version:
+                return self._dataset_id_cache
+
+        state = self._client.get_dataset_state()
+        payload = {
+            "version": self._dataset_version,
+            "frame": state.get("frame"),
+            "n": state.get("n"),
+            "k": state.get("k"),
+            "sortlist": state.get("sortlist"),
+        }
+        digest = _stable_hash(payload)
+
+        with self._lock:
+            self._dataset_id_cache = digest
+            self._dataset_id_cache_at_version = self._dataset_version
+            return digest
+
+    def get_view(self, view_id: str) -> Optional[ViewHandle]:
+        now = time.time()
+        with self._lock:
+            self._evict_expired_locked(now)
+            view = self._views.get(view_id)
+            if view is None:
+                return None
+            view.last_access = now
+            return view
+
+    def create_view(self, *, dataset_id: str, frame: str, filter_expr: str) -> ViewHandle:
+        current_id = self.current_dataset_id()
+        if dataset_id != current_id:
+            raise DatasetChangedError(current_id)
+
+        try:
+            obs_indices = self._client.compute_view_indices(filter_expr)
+        except ValueError as e:
+            raise InvalidFilterError(str(e))
+        except RuntimeError as e:
+            msg = str(e) or "No data in memory"
+            if "no data" in msg.lower():
+                raise NoDataInMemoryError(msg)
+            raise
+        now = time.time()
+        view_id = f"view_{uuid.uuid4().hex}"
+        view = ViewHandle(
+            view_id=view_id,
+            dataset_id=current_id,
+            frame=frame,
+            filter_expr=filter_expr,
+            obs_indices=obs_indices,
+            filtered_n=len(obs_indices),
+            created_at=now,
+            last_access=now,
+        )
+        with self._lock:
+            self._evict_expired_locked(now)
+            self._views[view_id] = view
+        return view
+
+    def delete_view(self, view_id: str) -> bool:
+        with self._lock:
+            return self._views.pop(view_id, None) is not None
+
+    def validate_token(self, header_value: str | None) -> bool:
+        if not header_value:
+            return False
+        if not header_value.startswith("Bearer "):
+            return False
+        token = header_value[len("Bearer ") :].strip()
+        with self._lock:
+            self._ensure_token()
+            if self._token is None:
+                return False
+            if time.time() * 1000 >= self._expires_at:
+                return False
+            return secrets.compare_digest(token, self._token)
+
+    def limits(self) -> tuple[int, int, int, int]:
+        return self._max_limit, self._max_vars, self._max_chars, self._max_request_bytes
+
+    def _ensure_token(self) -> None:
+        now_ms = int(time.time() * 1000)
+        if self._token is None or now_ms >= self._expires_at:
+            self._token = secrets.token_urlsafe(32)
+            self._expires_at = int((time.time() + self._token_ttl_s) * 1000)
+
+    def _evict_expired_locked(self, now: float) -> None:
+        expired: list[str] = []
+        for key, view in self._views.items():
+            if now - view.last_access >= self._view_ttl_s:
+                expired.append(key)
+        for key in expired:
+            self._views.pop(key, None)
+
+    def _ensure_http_server(self) -> None:
+        with self._lock:
+            if self._httpd is not None:
+                return
+
+            manager = self
+
+            class Handler(BaseHTTPRequestHandler):
+
+                def _send_json(self, status: int, payload: dict[str, Any]) -> None:
+                    data = json.dumps(payload).encode("utf-8")
+                    self.send_response(status)
+                    self.send_header("Content-Type", "application/json")
+                    self.send_header("Content-Length", str(len(data)))
+                    self.end_headers()
+                    self.wfile.write(data)
+
+                def _send_binary(self, status: int, data: bytes, content_type: str) -> None:
+                    self.send_response(status)
+                    self.send_header("Content-Type", content_type)
+                    self.send_header("Content-Length", str(len(data)))
+                    self.end_headers()
+                    self.wfile.write(data)
+
+                def _error(self, status: int, code: str, message: str, *, stata_rc: int | None = None) -> None:
+                    if status >= 500 or code == "internal_error":
+                        logger.error("UI HTTP error %s: %s", code, message)
+                        message = "Internal server error"
+                    body: dict[str, Any] = {"error": {"code": code, "message": message}}
+                    if stata_rc is not None:
+                        body["error"]["stataRc"] = stata_rc
+                    self._send_json(status, body)
+
+                def _require_auth(self) -> bool:
+                    if manager.validate_token(self.headers.get("Authorization")):
+                        return True
+                    self._error(401, "auth_failed", "Unauthorized")
+                    return False
+
+                def _read_json(self) -> dict[str, Any] | None:
+                    max_limit, max_vars, max_chars, max_bytes = manager.limits()
+                    _ = (max_limit, max_vars, max_chars)
+
+                    length = int(self.headers.get("Content-Length", "0") or "0")
+                    if length <= 0:
+                        return {}
+                    if length > max_bytes:
+                        self._error(400, "request_too_large", "Request too large")
+                        return None
+                    raw = self.rfile.read(length)
+                    try:
+                        parsed = json.loads(raw.decode("utf-8"))
+                    except Exception:
+                        self._error(400, "invalid_request", "Invalid JSON")
+                        return None
+                    if not isinstance(parsed, dict):
+                        self._error(400, "invalid_request", "Expected JSON object")
+                        return None
+                    return parsed
+
+                def do_GET(self) -> None:
+                    if not self._require_auth():
+                        return
+
+                    if self.path == "/v1/dataset":
+                        try:
+                            state = manager._client.get_dataset_state()
+                            dataset_id = manager.current_dataset_id()
+                            self._send_json(
+                                200,
+                                {
+                                    "dataset": {
+                                        "id": dataset_id,
+                                        "frame": state.get("frame"),
+                                        "n": state.get("n"),
+                                        "k": state.get("k"),
+                                        "changed": state.get("changed"),
+                                    }
+                                },
+                            )
+                            return
+                        except NoDataInMemoryError as e:
+                            self._error(400, "no_data_in_memory", str(e), stata_rc=e.stata_rc)
+                            return
+                        except Exception as e:
+                            self._error(500, "internal_error", str(e))
+                            return
+
+                    if self.path == "/v1/vars":
+                        try:
+                            state = manager._client.get_dataset_state()
+                            dataset_id = manager.current_dataset_id()
+                            variables = manager._client.list_variables_rich()
+                            self._send_json(
+                                200,
+                                {
+                                    "dataset": {"id": dataset_id, "frame": state.get("frame")},
+                                    "variables": variables,
+                                },
+                            )
+                            return
+                        except NoDataInMemoryError as e:
+                            self._error(400, "no_data_in_memory", str(e), stata_rc=e.stata_rc)
+                            return
+                        except Exception as e:
+                            self._error(500, "internal_error", str(e))
+                            return
+
+                    self._error(404, "not_found", "Not found")
+
+                def do_POST(self) -> None:
+                    if not self._require_auth():
+                        return
+
+
+                    if self.path == "/v1/arrow":
+                        body = self._read_json()
+                        if body is None:
+                            return
+                        try:
+                            resp_bytes = handle_arrow_request(manager, body, view_id=None)
+                            self._send_binary(200, resp_bytes, "application/vnd.apache.arrow.stream")
+                            return
+                        except HTTPError as e:
+                            self._error(e.status, e.code, e.message, stata_rc=e.stata_rc)
+                            return
+                        except Exception as e:
+                            self._error(500, "internal_error", str(e))
+                            return
+
+                    if self.path == "/v1/page":
+                        body = self._read_json()
+                        if body is None:
+                            return
+                        try:
+                            resp = handle_page_request(manager, body, view_id=None)
+                            self._send_json(200, resp)
+                            return
+                        except HTTPError as e:
+                            self._error(e.status, e.code, e.message, stata_rc=e.stata_rc)
+                            return
+                        except Exception as e:
+                            self._error(500, "internal_error", str(e))
+                            return
+
+                    if self.path == "/v1/views":
+                        body = self._read_json()
+                        if body is None:
+                            return
+                        dataset_id = str(body.get("datasetId", ""))
+                        frame = str(body.get("frame", "default"))
+                        filter_expr = str(body.get("filterExpr", ""))
+                        if not dataset_id or not filter_expr:
+                            self._error(400, "invalid_request", "datasetId and filterExpr are required")
+                            return
+                        try:
+                            view = manager.create_view(dataset_id=dataset_id, frame=frame, filter_expr=filter_expr)
+                            self._send_json(
+                                200,
+                                {
+                                    "dataset": {"id": view.dataset_id, "frame": view.frame},
+                                    "view": {"id": view.view_id, "filteredN": view.filtered_n},
+                                },
+                            )
+                            return
+                        except DatasetChangedError as e:
+                            self._error(409, "dataset_changed", "Dataset changed")
+                            return
+                        except ValueError as e:
+                            self._error(400, "invalid_filter", str(e))
+                            return
+                        except RuntimeError as e:
+                            msg = str(e) or "No data in memory"
+                            if "no data" in msg.lower():
+                                self._error(400, "no_data_in_memory", msg)
+                                return
+                            self._error(500, "internal_error", msg)
+                            return
+                        except Exception as e:
+                            self._error(500, "internal_error", str(e))
+                            return
+
+                    if self.path.startswith("/v1/views/") and self.path.endswith("/page"):
+                        parts = self.path.split("/")
+                        if len(parts) != 5:
+                            self._error(404, "not_found", "Not found")
+                            return
+                        view_id = parts[3]
+                        body = self._read_json()
+                        if body is None:
+                            return
+                        try:
+                            resp = handle_page_request(manager, body, view_id=view_id)
+                            self._send_json(200, resp)
+                            return
+                        except HTTPError as e:
+                            self._error(e.status, e.code, e.message, stata_rc=e.stata_rc)
+                            return
+                        except Exception as e:
+                            self._error(500, "internal_error", str(e))
+                            return
+
+                    if self.path.startswith("/v1/views/") and self.path.endswith("/arrow"):
+                        parts = self.path.split("/")
+                        if len(parts) != 5:
+                            self._error(404, "not_found", "Not found")
+                            return
+                        view_id = parts[3]
+                        body = self._read_json()
+                        if body is None:
+                            return
+                        try:
+                            resp_bytes = handle_arrow_request(manager, body, view_id=view_id)
+                            self._send_binary(200, resp_bytes, "application/vnd.apache.arrow.stream")
+                            return
+                        except HTTPError as e:
+                            self._error(e.status, e.code, e.message, stata_rc=e.stata_rc)
+                            return
+                        except Exception as e:
+                            self._error(500, "internal_error", str(e))
+                            return
+
+                    if self.path == "/v1/filters/validate":
+                        body = self._read_json()
+                        if body is None:
+                            return
+                        filter_expr = str(body.get("filterExpr", ""))
+                        if not filter_expr:
+                            self._error(400, "invalid_request", "filterExpr is required")
+                            return
+                        try:
+                            manager._client.validate_filter_expr(filter_expr)
+                            self._send_json(200, {"ok": True})
+                            return
+                        except ValueError as e:
+                            self._error(400, "invalid_filter", str(e))
+                            return
+                        except RuntimeError as e:
+                            msg = str(e) or "No data in memory"
+                            if "no data" in msg.lower():
+                                self._error(400, "no_data_in_memory", msg)
+                                return
+                            self._error(500, "internal_error", msg)
+                            return
+                        except Exception as e:
+                            self._error(500, "internal_error", str(e))
+                            return
+
+                    self._error(404, "not_found", "Not found")
+
+                def do_DELETE(self) -> None:
+                    if not self._require_auth():
+                        return
+
+                    if self.path.startswith("/v1/views/"):
+                        parts = self.path.split("/")
+                        if len(parts) != 4:
+                            self._error(404, "not_found", "Not found")
+                            return
+                        view_id = parts[3]
+                        if manager.delete_view(view_id):
+                            self._send_json(200, {"ok": True})
+                        else:
+                            self._error(404, "not_found", "Not found")
+                        return
+
+                    self._error(404, "not_found", "Not found")
+
+                def log_message(self, format: str, *args: Any) -> None:
+                    return
+
+            httpd = ThreadingHTTPServer((self._host, self._port), Handler)
+            t = threading.Thread(target=httpd.serve_forever, daemon=True)
+            t.start()
+            self._httpd = httpd
+            self._thread = t
+
+
+class HTTPError(Exception):
+    def __init__(self, status: int, code: str, message: str, *, stata_rc: int | None = None):
+        super().__init__(message)
+        self.status = status
+        self.code = code
+        self.message = message
+        self.stata_rc = stata_rc
+
+
+class DatasetChangedError(Exception):
+    def __init__(self, current_dataset_id: str):
+        super().__init__("dataset_changed")
+        self.current_dataset_id = current_dataset_id
+
+
+class NoDataInMemoryError(Exception):
+    def __init__(self, message: str = "No data in memory", *, stata_rc: int | None = None):
+        super().__init__(message)
+        self.stata_rc = stata_rc
+
+
+class InvalidFilterError(Exception):
+    def __init__(self, message: str, *, stata_rc: int | None = None):
+        super().__init__(message)
+        self.message = message
+        self.stata_rc = stata_rc
+
+
+def handle_page_request(manager: UIChannelManager, body: dict[str, Any], *, view_id: str | None) -> dict[str, Any]:
+    max_limit, max_vars, max_chars, _ = manager.limits()
+
+    if view_id is None:
+        dataset_id = str(body.get("datasetId", ""))
+        frame = str(body.get("frame", "default"))
+    else:
+        view = manager.get_view(view_id)
+        if view is None:
+            raise HTTPError(404, "not_found", "View not found")
+        dataset_id = view.dataset_id
+        frame = view.frame
+
+    # Parse offset (default 0 is valid since offset >= 0)
+    try:
+        offset = int(body.get("offset") or 0)
+    except (ValueError, TypeError) as e:
+        raise HTTPError(400, "invalid_request", f"offset must be a valid integer, got: {body.get('offset')!r}")
+
+    # Parse limit (no default - must be explicitly provided)
+    limit_raw = body.get("limit")
+    if limit_raw is None:
+        raise HTTPError(400, "invalid_request", "limit is required")
+    try:
+        limit = int(limit_raw)
+    except (ValueError, TypeError) as e:
+        raise HTTPError(400, "invalid_request", f"limit must be a valid integer, got: {limit_raw!r}")
+
+    vars_req = body.get("vars", [])
+    include_obs_no = bool(body.get("includeObsNo", False))
+
+    # Parse sortBy parameter
+    sort_by = body.get("sortBy", [])
+    if sort_by is not None and not isinstance(sort_by, list):
+        raise HTTPError(400, "invalid_request", f"sortBy must be an array, got: {type(sort_by).__name__}")
+    if sort_by and not all(isinstance(s, str) for s in sort_by):
+        raise HTTPError(400, "invalid_request", "sortBy must be an array of strings")
+
+    # Parse maxChars
+    max_chars_raw = body.get("maxChars", max_chars)
+    try:
+        max_chars_req = int(max_chars_raw or max_chars)
+    except (ValueError, TypeError) as e:
+        raise HTTPError(400, "invalid_request", f"maxChars must be a valid integer, got: {max_chars_raw!r}")
+
+    if offset < 0:
+        raise HTTPError(400, "invalid_request", f"offset must be >= 0, got: {offset}")
+    if limit <= 0:
+        raise HTTPError(400, "invalid_request", f"limit must be > 0, got: {limit}")
+    if limit > max_limit:
+        raise HTTPError(400, "request_too_large", f"limit must be <= {max_limit}")
+    if max_chars_req <= 0:
+        raise HTTPError(400, "invalid_request", "maxChars must be > 0")
+    if max_chars_req > max_chars:
+        raise HTTPError(400, "request_too_large", f"maxChars must be <= {max_chars}")
+
+    if not isinstance(vars_req, list) or not all(isinstance(v, str) for v in vars_req):
+        raise HTTPError(400, "invalid_request", "vars must be a list of strings")
+    if len(vars_req) > max_vars:
+        raise HTTPError(400, "request_too_large", f"vars length must be <= {max_vars}")
+
+    current_id = manager.current_dataset_id()
+    if dataset_id != current_id:
+        raise HTTPError(409, "dataset_changed", "Dataset changed")
+
+    if view_id is None:
+        obs_indices = None
+        filtered_n: int | None = None
+    else:
+        assert view is not None
+        obs_indices = view.obs_indices
+        filtered_n = view.filtered_n
+
+    try:
+        # Apply sorting if requested (Rust native sorter with Polars fallback; no dataset mutation)
+        if sort_by:
+            try:
+                normalized_sort = manager._normalize_sort_spec(sort_by)
+                sorted_indices = manager._get_cached_sort_indices(current_id, normalized_sort)
+                if sorted_indices is None:
+                    sort_cols = [spec.lstrip("+-") for spec in normalized_sort]
+                    descending = [spec.startswith("-") for spec in normalized_sort]
+                    nulls_last = [not desc for desc in descending]
+
+                    table = manager._get_sort_table(current_id, sort_cols)
+                    if table is None:
+                        sorted_indices = []
+                    else:
+                        sorted_indices = _try_native_argsort(table, sort_cols, descending, nulls_last)
+                        if sorted_indices is None:
+                            sorted_indices = _get_sorted_indices_polars(table, sort_cols, descending, nulls_last)
+
+                    manager._set_cached_sort_indices(current_id, normalized_sort, sorted_indices)
+
+                if view_id is None:
+                    obs_indices = sorted_indices
+                else:
+                    assert view is not None
+                    view_set = set(view.obs_indices)
+                    obs_indices = [idx for idx in sorted_indices if idx in view_set]
+                    filtered_n = len(obs_indices)
+            except ValueError as e:
+                raise HTTPError(400, "invalid_request", f"Invalid sort specification: {e}")
+            except RuntimeError as e:
+                raise HTTPError(500, "internal_error", f"Failed to apply sort: {e}")
+
+        dataset_state = manager._client.get_dataset_state()
+        page = manager._client.get_page(
+            offset=offset,
+            limit=limit,
+            vars=vars_req,
+            include_obs_no=include_obs_no,
+            max_chars=max_chars_req,
+            obs_indices=obs_indices,
+        )
+    except HTTPError:
+        # Re-raise HTTPError exceptions as-is
+        raise
+    except RuntimeError as e:
+        msg = str(e) or "No data in memory"
+        if "no data" in msg.lower():
+            raise HTTPError(400, "no_data_in_memory", msg)
+        raise HTTPError(500, "internal_error", msg)
+    except ValueError as e:
+        msg = str(e)
+        if msg.lower().startswith("invalid variable"):
+            raise HTTPError(400, "invalid_variable", msg)
+        raise HTTPError(400, "invalid_request", msg)
+    except Exception as e:
+        raise HTTPError(500, "internal_error", str(e))
+
+    view_obj: dict[str, Any] = {
+        "offset": offset,
+        "limit": limit,
+        "returned": page["returned"],
+        "filteredN": filtered_n,
+    }
+    if view_id is not None:
+        view_obj["viewId"] = view_id
+
+    return {
+        "dataset": {
+            "id": current_id,
+            "frame": dataset_state.get("frame"),
+            "n": dataset_state.get("n"),
+            "k": dataset_state.get("k"),
+        },
+        "view": view_obj,
+        "vars": page["vars"],
+        "rows": page["rows"],
+        "display": {
+            "maxChars": max_chars_req,
+            "truncatedCells": page["truncated_cells"],
+            "missing": ".",
+        },
+    }
+
+
+def handle_arrow_request(manager: UIChannelManager, body: dict[str, Any], *, view_id: str | None) -> bytes:
+    max_limit, max_vars, max_chars, _ = manager.limits()
+    # Use the specific Arrow limit instead of the general UI page limit
+    chunk_limit = getattr(manager, "_max_arrow_limit", 1_000_000)
+
+    if view_id is None:
+        dataset_id = str(body.get("datasetId", ""))
+        frame = str(body.get("frame", "default"))
+    else:
+        view = manager.get_view(view_id)
+        if view is None:
+            raise HTTPError(404, "not_found", "View not found")
+        dataset_id = view.dataset_id
+        frame = view.frame
+
+    # Parse offset (default 0)
+    try:
+        offset = int(body.get("offset") or 0)
+    except (ValueError, TypeError):
+        raise HTTPError(400, "invalid_request", "offset must be a valid integer")
+
+    # Parse limit (required)
+    limit_raw = body.get("limit")
+    if limit_raw is None:
+        # Default to the max arrow limit if not specified? 
+        # The previous code required it. Let's keep it required but allow large values.
+        raise HTTPError(400, "invalid_request", "limit is required")
+    try:
+        limit = int(limit_raw)
+    except (ValueError, TypeError):
+        raise HTTPError(400, "invalid_request", "limit must be a valid integer")
+
+    vars_req = body.get("vars", [])
+    include_obs_no = bool(body.get("includeObsNo", False))
+    sort_by = body.get("sortBy", [])
+
+    if offset < 0:
+        raise HTTPError(400, "invalid_request", "offset must be >= 0")
+    if limit <= 0:
+        raise HTTPError(400, "invalid_request", "limit must be > 0")
+    # Arrow streams are efficient, but we still respect a (much larger) max limit
+    if limit > chunk_limit:
+        raise HTTPError(400, "request_too_large", f"limit must be <= {chunk_limit}")
+
+    if not isinstance(vars_req, list) or not all(isinstance(v, str) for v in vars_req):
+        raise HTTPError(400, "invalid_request", "vars must be a list of strings")
+    if len(vars_req) > max_vars:
+        raise HTTPError(400, "request_too_large", f"vars length must be <= {max_vars}")
+
+    current_id = manager.current_dataset_id()
+    if dataset_id != current_id:
+        raise HTTPError(409, "dataset_changed", "Dataset changed")
+
+    if view_id is None:
+        obs_indices = None
+    else:
+        assert view is not None
+        obs_indices = view.obs_indices
+
+    try:
+        # Apply sorting if requested (Rust native sorter with Polars fallback; no dataset mutation)
+        if sort_by:
+            if not isinstance(sort_by, list) or not all(isinstance(s, str) for s in sort_by):
+                raise HTTPError(400, "invalid_request", "sortBy must be a list of strings")
+            try:
+                normalized_sort = manager._normalize_sort_spec(sort_by)
+                sorted_indices = manager._get_cached_sort_indices(current_id, normalized_sort)
+                if sorted_indices is None:
+                    sort_cols = [spec.lstrip("+-") for spec in normalized_sort]
+                    descending = [spec.startswith("-") for spec in normalized_sort]
+                    nulls_last = [not desc for desc in descending]
+
+                    table = manager._get_sort_table(current_id, sort_cols)
+                    if table is None:
+                        sorted_indices = []
+                    else:
+                        sorted_indices = _try_native_argsort(table, sort_cols, descending, nulls_last)
+                        if sorted_indices is None:
+                            sorted_indices = _get_sorted_indices_polars(table, sort_cols, descending, nulls_last)
+
+                manager._set_cached_sort_indices(current_id, normalized_sort, sorted_indices)
+
+                if view_id is None:
+                    obs_indices = sorted_indices
+                else:
+                    assert view is not None
+                    view_set = set(view.obs_indices)
+                    obs_indices = [idx for idx in sorted_indices if idx in view_set]
+            except ValueError as e:
+                raise HTTPError(400, "invalid_request", f"Invalid sort: {e}")
+            except RuntimeError as e:
+                raise HTTPError(500, "internal_error", f"Sort failed: {e}")
+
+        arrow_bytes = manager._client.get_arrow_stream(
+            offset=offset,
+            limit=limit,
+            vars=vars_req,
+            include_obs_no=include_obs_no,
+            obs_indices=obs_indices,
+        )
+        return arrow_bytes
+
+    except RuntimeError as e:
+        msg = str(e) or "No data in memory"
+        if "no data" in msg.lower():
+            raise HTTPError(400, "no_data_in_memory", msg)
+        raise HTTPError(500, "internal_error", msg)
+    except ValueError as e:
+        msg = str(e)
+        if "invalid variable" in msg.lower():
+            raise HTTPError(400, "invalid_variable", msg)
+        raise HTTPError(400, "invalid_request", msg)
+    except Exception as e:
+        raise HTTPError(500, "internal_error", str(e))
+