fix(node): reject non-ascii cfg keys

nodejs/.cargo/config.toml

@@ -0,0 +1,21 @@
+# Keep Node's Linux x64 addons compatible with pre-Haswell CPUs.
+# lance-linalg dispatches hot vector kernels to newer SIMD tiers at runtime.
+[env]
+LANCEDB_NODE_ENFORCE_X86_64_V2 = "1"
+
+[target.x86_64-unknown-linux-gnu]
+rustflags = [
+    "-C",
+    "target-cpu=x86-64-v2",
+    "-C",
+    "target-feature=-avx,-avx2,-fma,-f16c",
+]
+
+# Preserve the workspace's dynamic C runtime configuration for musl.
+[target.x86_64-unknown-linux-musl]
+rustflags = [
+    "-C",
+    "target-cpu=x86-64-v2",
+    "-C",
+    "target-feature=-crt-static,-avx,-avx2,-fma,-f16c",
+]

nodejs/Cargo.toml

@@ -12,6 +12,10 @@ categories.workspace = true
 [lib]
 crate-type = ["cdylib"]
 
+[[test]]
+name = "x86_64_v2_build_flags"
+path = "build_support/x86_64_v2.rs"
+
 [dependencies]
 async-trait.workspace = true
 arrow-ipc.workspace = true

nodejs/__test__/cpu_compatibility.test.ts

@@ -0,0 +1,50 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+import * as tmp from "tmp";
+
+import { connect } from "../lancedb";
+import {
+  Field,
+  FixedSizeList,
+  Float32,
+  Int32,
+  Schema,
+  makeArrowTable,
+} from "../lancedb/arrow";
+
+test("cosine vector search runs on the pre-Haswell build baseline", async () => {
+  const tmpDir = tmp.dirSync({ unsafeCleanup: true });
+
+  try {
+    const db = await connect(tmpDir.name);
+    const schema = new Schema([
+      new Field("id", new Int32(), false),
+      new Field(
+        "vector",
+        new FixedSizeList(3, new Field("item", new Float32(), false)),
+        false,
+      ),
+    ]);
+    const data = makeArrowTable(
+      [
+        { id: 1, vector: [1, 0, 0] },
+        { id: 2, vector: [0, 1, 0] },
+      ],
+      { schema },
+    );
+    const table = await db.createTable("vectors", data);
+
+    const results = await table
+      .vectorSearch([1, 0, 0])
+      .distanceType("cosine")
+      .limit(1)
+      .toArray();
+
+    expect(results).toHaveLength(1);
+    expect(results[0].id).toBe(1);
+    expect(results[0]._distance).toBeCloseTo(0);
+  } finally {
+    tmpDir.removeCallback();
+  }
+}, 60_000);

nodejs/build.rs

@@ -3,6 +3,29 @@
 
 extern crate napi_build;
 
+use std::env;
+
+#[path = "build_support/x86_64_v2.rs"]
+mod x86_64_v2;
+
+const ENFORCE_BASELINE: &str = "LANCEDB_NODE_ENFORCE_X86_64_V2";
+
 fn main() {
     napi_build::setup();
+    println!("cargo:rerun-if-env-changed={ENFORCE_BASELINE}");
+    println!("cargo:rerun-if-env-changed=CARGO_ENCODED_RUSTFLAGS");
+
+    let is_linux_x64 = env::var("CARGO_CFG_TARGET_ARCH").as_deref() == Ok("x86_64")
+        && env::var("CARGO_CFG_TARGET_OS").as_deref() == Ok("linux");
+    let is_release = env::var("PROFILE").as_deref() == Ok("release");
+    let is_node_build = env::var(ENFORCE_BASELINE).as_deref() == Ok("1");
+
+    if !is_linux_x64 || (!is_release && !is_node_build) {
+        return;
+    }
+
+    let encoded_rustflags = env::var("CARGO_ENCODED_RUSTFLAGS").unwrap_or_default();
+    x86_64_v2::validate_encoded_rustflags(&encoded_rustflags).unwrap_or_else(|error| {
+        panic!("Linux x64 Node addons must use the x86-64-v2 baseline; {error}")
+    });
 }

nodejs/build_support/x86_64_v2.rs

@@ -0,0 +1,430 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+use std::collections::{BTreeMap, BTreeSet};
+
+const SAFE_TARGET_CPUS: [&str; 2] = ["x86-64", "x86-64-v2"];
+const BASELINE_FEATURES: [&str; 9] = [
+    "cmpxchg16b",
+    "fxsr",
+    "popcnt",
+    "sse",
+    "sse2",
+    "sse3",
+    "sse4.1",
+    "sse4.2",
+    "ssse3",
+];
+
+pub(crate) fn validate_encoded_rustflags(encoded: &str) -> Result<(), String> {
+    let mut target_cpu = None;
+    let mut feature_states = BTreeMap::new();
+    let mut required_disables = BTreeSet::new();
+    let mut unsupported_features = BTreeSet::new();
+
+    for option in codegen_options(encoded)? {
+        let Some((name, value)) = option.split_once('=') else {
+            continue;
+        };
+
+        let name = name.replace('_', "-");
+        match name.as_str() {
+            "target-cpu" => target_cpu = Some(value),
+            "target-feature" => {
+                for toggle in value.split(',').filter(|toggle| !toggle.is_empty()) {
+                    let (enabled, feature) = match toggle.as_bytes()[0] {
+                        b'+' => (true, &toggle[1..]),
+                        b'-' => (false, &toggle[1..]),
+                        _ => return Err(format!("invalid target feature flag: {toggle}")),
+                    };
+
+                    feature_states.insert(feature, enabled);
+                    if enabled && !BASELINE_FEATURES.contains(&feature) {
+                        match feature {
+                            // These are inherited from the workspace configuration and
+                            // explicitly canceled by the Node configuration. Account for
+                            // their implied AVX prerequisite as well as the named feature.
+                            "avx" => {
+                                required_disables.insert("avx");
+                            }
+                            "avx2" => {
+                                required_disables.extend(["avx", "avx2"]);
+                            }
+                            "f16c" => {
+                                required_disables.extend(["avx", "f16c"]);
+                            }
+                            "fma" => {
+                                required_disables.extend(["avx", "fma"]);
+                            }
+                            _ => {
+                                unsupported_features.insert(feature);
+                            }
+                        }
+                    }
+                }
+            }
+            // LLVM arguments can independently alter the target feature set and
+            // cannot be proven safe by inspecting rustc's target options.
+            "llvm-args" => return Err("LLVM arguments can override the CPU baseline".to_owned()),
+            _ => {}
+        }
+    }
+
+    if let Some(cpu) = target_cpu.filter(|cpu| !SAFE_TARGET_CPUS.contains(cpu)) {
+        return Err(format!(
+            "effective target CPU is {}, expected x86-64-v2 or lower",
+            cpu
+        ));
+    }
+
+    if !unsupported_features.is_empty() {
+        return Err(format!(
+            "features above v2: {}",
+            unsupported_features
+                .into_iter()
+                .collect::<Vec<_>>()
+                .join(", ")
+        ));
+    }
+
+    let not_disabled = required_disables
+        .into_iter()
+        .filter(|feature| feature_states.get(feature) != Some(&false))
+        .collect::<Vec<_>>();
+    if !not_disabled.is_empty() {
+        return Err(format!(
+            "inherited features not fully disabled: {}",
+            not_disabled.join(", ")
+        ));
+    }
+
+    Ok(())
+}
+
+fn codegen_options(encoded: &str) -> Result<Vec<&str>, String> {
+    let arguments = encoded.split('\u{1f}').collect::<Vec<_>>();
+    if arguments.iter().any(|argument| argument.starts_with('@')) {
+        return Err("rustc response-file arguments cannot be validated".to_owned());
+    }
+
+    let mut options = Vec::new();
+    let mut index = 0;
+
+    while index < arguments.len() {
+        let argument = arguments[index];
+        if argument == "--cfg" {
+            index += 1;
+            let cfg = arguments
+                .get(index)
+                .copied()
+                .ok_or_else(|| "missing value after --cfg".to_owned())?;
+            reject_builtin_target_feature_cfg(cfg)?;
+        } else if let Some(cfg) = argument.strip_prefix("--cfg=") {
+            reject_builtin_target_feature_cfg(cfg)?;
+        } else if argument == "-C" || argument == "--codegen" {
+            index += 1;
+            let option = arguments
+                .get(index)
+                .copied()
+                .ok_or_else(|| format!("missing value after {argument}"))?;
+            options.push(option.trim_start_matches('='));
+        } else if let Some(option) = argument.strip_prefix("-C") {
+            if !option.is_empty() {
+                options.push(option.trim_start_matches('='));
+            }
+        } else if let Some(option) = argument.strip_prefix("--codegen=") {
+            options.push(option);
+        }
+        index += 1;
+    }
+
+    Ok(options)
+}
+
+fn reject_builtin_target_feature_cfg(cfg: &str) -> Result<(), String> {
+    let key = cfg.split_once('=').map_or(cfg, |(key, _)| key);
+    if key.contains("/*") || key.contains("//") {
+        return Err("comment-bearing cfgs cannot be validated".to_owned());
+    }
+    if !key.is_ascii() {
+        return Err("non-ASCII cfg keys cannot be validated".to_owned());
+    }
+
+    let name = key.trim();
+    let name = name.strip_prefix("r#").unwrap_or(name);
+    if name == "target_feature" {
+        return Err("built-in target_feature cfgs can override runtime CPU detection".to_owned());
+    }
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn encoded(arguments: &[&str]) -> String {
+        arguments.join("\u{1f}")
+    }
+
+    #[test]
+    fn accepts_merged_workspace_and_node_flags() {
+        let flags = encoded(&[
+            "-C",
+            "target-cpu=haswell",
+            "-C",
+            "target-feature=+avx2,+fma,+f16c",
+            "-C",
+            "target-cpu=x86-64-v2",
+            "-C",
+            "target-feature=-avx,-avx2,-fma,-f16c",
+        ]);
+
+        assert_eq!(validate_encoded_rustflags(&flags), Ok(()));
+    }
+
+    #[test]
+    fn accepts_non_merging_v2_boundary() {
+        let flags = encoded(&["-Ctarget-cpu=x86-64-v2"]);
+
+        assert_eq!(validate_encoded_rustflags(&flags), Ok(()));
+    }
+
+    #[test]
+    fn accepts_default_cpu_with_non_codegen_flags() {
+        let flags = encoded(&["-D", "warnings"]);
+
+        assert_eq!(validate_encoded_rustflags(&flags), Ok(()));
+    }
+
+    #[test]
+    fn accepts_unrelated_custom_cfg() {
+        let flags = encoded(&["--cfg=tokio_unstable"]);
+
+        assert_eq!(validate_encoded_rustflags(&flags), Ok(()));
+    }
+
+    #[test]
+    fn accepts_comment_like_syntax_in_cfg_value() {
+        let flags = encoded(&[r#"--cfg=endpoint="https://example.com/*""#]);
+
+        assert_eq!(validate_encoded_rustflags(&flags), Ok(()));
+    }
+
+    #[test]
+    fn accepts_explicit_x86_64_v1_cpu() {
+        let flags = encoded(&["-Ctarget-cpu=x86-64"]);
+
+        assert_eq!(validate_encoded_rustflags(&flags), Ok(()));
+    }
+
+    #[test]
+    fn accepts_musl_dynamic_crt_configuration() {
+        let flags = encoded(&[
+            "-C",
+            "target-cpu=haswell",
+            "-C",
+            "target-feature=-crt-static,+avx2,+fma,+f16c",
+            "-C",
+            "target-cpu=x86-64-v2",
+            "-C",
+            "target-feature=-crt-static,-avx,-avx2,-fma,-f16c",
+        ]);
+
+        assert_eq!(validate_encoded_rustflags(&flags), Ok(()));
+    }
+
+    #[test]
+    fn rejects_feature_omitted_from_target_cfg() {
+        let flags = encoded(&["-Ctarget-cpu=x86-64-v2", "-Ctarget-feature=+apxf"]);
+
+        assert_eq!(
+            validate_encoded_rustflags(&flags),
+            Err("features above v2: apxf".to_owned())
+        );
+    }
+
+    #[test]
+    fn rejects_underscore_spelling_above_baseline_feature() {
+        let flags = encoded(&["-Ctarget_cpu=x86-64-v2", "-Ctarget_feature=+apxf"]);
+
+        assert_eq!(
+            validate_encoded_rustflags(&flags),
+            Err("features above v2: apxf".to_owned())
+        );
+    }
+
+    #[test]
+    fn rejects_unexpected_above_baseline_feature() {
+        let flags = encoded(&[
+            "--codegen=target-cpu=x86-64-v2",
+            "--codegen",
+            "target-feature=+bmi2",
+        ]);
+
+        assert_eq!(
+            validate_encoded_rustflags(&flags),
+            Err("features above v2: bmi2".to_owned())
+        );
+    }
+
+    #[test]
+    fn rejects_cpu_above_v2() {
+        let flags = encoded(&["-Ctarget-cpu=haswell"]);
+
+        assert_eq!(
+            validate_encoded_rustflags(&flags),
+            Err("effective target CPU is haswell, expected x86-64-v2 or lower".to_owned())
+        );
+    }
+
+    #[test]
+    fn rejects_incompletely_disabled_feature_implications() {
+        let flags = encoded(&["-Ctarget-cpu=x86-64-v2", "-Ctarget-feature=+avx2,-avx2"]);
+
+        assert_eq!(
+            validate_encoded_rustflags(&flags),
+            Err("inherited features not fully disabled: avx".to_owned())
+        );
+    }
+
+    #[test]
+    fn rejects_llvm_feature_overrides() {
+        let flags = encoded(&["-Ctarget-cpu=x86-64-v2", "-Cllvm-args=-mattr=+apxf"]);
+
+        assert_eq!(
+            validate_encoded_rustflags(&flags),
+            Err("LLVM arguments can override the CPU baseline".to_owned())
+        );
+    }
+
+    #[test]
+    fn rejects_underscore_spelling_llvm_feature_overrides() {
+        let flags = encoded(&["-Ctarget_cpu=x86-64-v2", "-Cllvm_args=-mattr=+apxf"]);
+
+        assert_eq!(
+            validate_encoded_rustflags(&flags),
+            Err("LLVM arguments can override the CPU baseline".to_owned())
+        );
+    }
+
+    #[test]
+    fn rejects_response_file_arguments() {
+        let flags = encoded(&["-Ctarget-cpu=x86-64-v2", "@flags.rsp"]);
+
+        assert_eq!(
+            validate_encoded_rustflags(&flags),
+            Err("rustc response-file arguments cannot be validated".to_owned())
+        );
+    }
+
+    #[test]
+    fn rejects_split_builtin_target_feature_cfg() {
+        let flags = encoded(&[
+            "-Ctarget-cpu=x86-64-v2",
+            "--cfg",
+            r#"target_feature="avx2""#,
+            "-Aexplicit_builtin_cfgs_in_flags",
+        ]);
+
+        assert_eq!(
+            validate_encoded_rustflags(&flags),
+            Err("built-in target_feature cfgs can override runtime CPU detection".to_owned())
+        );
+    }
+
+    #[test]
+    fn rejects_equals_builtin_target_feature_cfg() {
+        let flags = encoded(&[
+            "-Ctarget-cpu=x86-64-v2",
+            r#"--cfg=target_feature="avx2""#,
+            "-Aexplicit_builtin_cfgs_in_flags",
+        ]);
+
+        assert_eq!(
+            validate_encoded_rustflags(&flags),
+            Err("built-in target_feature cfgs can override runtime CPU detection".to_owned())
+        );
+    }
+
+    #[test]
+    fn rejects_raw_identifier_builtin_target_feature_cfg() {
+        let flags = encoded(&[
+            "-Ctarget-cpu=x86-64-v2",
+            r#"--cfg=r#target_feature="avx2""#,
+            "-Aexplicit_builtin_cfgs_in_flags",
+        ]);
+
+        assert_eq!(
+            validate_encoded_rustflags(&flags),
+            Err("built-in target_feature cfgs can override runtime CPU detection".to_owned())
+        );
+    }
+
+    #[test]
+    fn rejects_block_comment_after_cfg_name() {
+        let flags = encoded(&[
+            "--cfg",
+            r#"target_feature/*gate*/="avx2""#,
+            "-Aexplicit_builtin_cfgs_in_flags",
+        ]);
+
+        assert_eq!(
+            validate_encoded_rustflags(&flags),
+            Err("comment-bearing cfgs cannot be validated".to_owned())
+        );
+    }
+
+    #[test]
+    fn rejects_block_comment_before_cfg_name() {
+        let flags = encoded(&[
+            r#"--cfg=/*gate*/target_feature="avx2""#,
+            "-Aexplicit_builtin_cfgs_in_flags",
+        ]);
+
+        assert_eq!(
+            validate_encoded_rustflags(&flags),
+            Err("comment-bearing cfgs cannot be validated".to_owned())
+        );
+    }
+
+    #[test]
+    fn rejects_line_comment_cfg_trivia() {
+        let flags = encoded(&[
+            "--cfg",
+            "target_feature// gate\n=\"avx2\"",
+            "-Aexplicit_builtin_cfgs_in_flags",
+        ]);
+
+        assert_eq!(
+            validate_encoded_rustflags(&flags),
+            Err("comment-bearing cfgs cannot be validated".to_owned())
+        );
+    }
+
+    #[test]
+    fn rejects_leading_bom_in_cfg_key() {
+        let flags = encoded(&[
+            "--cfg",
+            "\u{feff}target_feature=\"avx2\"",
+            "-Aexplicit_builtin_cfgs_in_flags",
+        ]);
+
+        assert_eq!(
+            validate_encoded_rustflags(&flags),
+            Err("non-ASCII cfg keys cannot be validated".to_owned())
+        );
+    }
+
+    #[test]
+    fn rejects_non_ascii_pattern_whitespace_in_cfg_key() {
+        for whitespace in ['\u{200e}', '\u{200f}'] {
+            let cfg = format!("{whitespace}target_feature=\"avx2\"");
+            let flags = encoded(&["--cfg", &cfg, "-Aexplicit_builtin_cfgs_in_flags"]);
+
+            assert_eq!(
+                validate_encoded_rustflags(&flags),
+                Err("non-ASCII cfg keys cannot be validated".to_owned())
+            );
+        }
+    }
+}

nodejs/ci/pre_haswell_smoke.js

@@ -0,0 +1,37 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright The LanceDB Authors
+
+const assert = require("node:assert/strict");
+const fs = require("node:fs");
+const os = require("node:os");
+const path = require("node:path");
+
+const { connect } = require("../dist");
+
+async function main() {
+  const tmpDir = fs.mkdtempSync(path.join(os.tmpdir(), "lancedb-cpu-"));
+
+  try {
+    const db = await connect(tmpDir);
+    const table = await db.createTable("vectors", [
+      { id: 1, vector: [1, 0, 0] },
+      { id: 2, vector: [0, 1, 0] },
+    ]);
+    const results = await table
+      .vectorSearch([1, 0, 0])
+      .distanceType("cosine")
+      .limit(1)
+      .toArray();
+
+    assert.equal(results.length, 1);
+    assert.equal(results[0].id, 1);
+    assert.ok(Math.abs(results[0]._distance) < 1e-6);
+  } finally {
+    fs.rmSync(tmpDir, { force: true, recursive: true });
+  }
+}
+
+main().catch((error) => {
+  console.error(error);
+  process.exitCode = 1;
+});

nodejs/package.json

@@ -88,6 +88,7 @@
     "lint-fix": "biome check --write . && biome format --write .",
     "prepublishOnly": "napi prepublish -t npm",
     "test": "jest --verbose",
+    "test:pre-haswell": "node ci/pre_haswell_smoke.js",
     "integration": "S3_TEST=1 pnpm test",
     "universal": "napi universalize",
     "version": "napi version"

python/python/lancedb/_lancedb.pyi

@@ -269,7 +269,6 @@ class Table:
         mode: Literal["append", "overwrite"],
         progress: Optional[Any] = None,
         write_parallelism: Optional[int] = None,
-        on_nan_vectors: Optional[Literal["error", "keep"]] = None,
     ) -> AddResult: ...
     async def update(
         self, updates: Dict[str, str], where: Optional[str]

python/python/lancedb/db.py

@@ -333,9 +333,7 @@ class DBConnection(EnforceOverrides):
             schema that's specified.
         on_bad_vectors: str, default "error"
             What to do if any of the vectors are not the same size or contains NaNs.
-            One of "error", "drop", "fill", "null", or "keep". With "keep",
-            vectors containing NaNs are preserved, but vectors with the wrong
-            dimension still raise an error.
+            One of "error", "drop", "fill".
         fill_value: float
             The value to use when filling vectors. Only used if on_bad_vectors="fill".
         storage_options: dict, optional
@@ -1597,9 +1595,7 @@ class AsyncConnection(object):
             schema that's specified.
         on_bad_vectors: str, default "error"
             What to do if any of the vectors are not the same size or contains NaNs.
-            One of "error", "drop", "fill", "null", or "keep". With "keep",
-            vectors containing NaNs are preserved, but vectors with the wrong
-            dimension still raise an error.
+            One of "error", "drop", "fill".
         fill_value: float
             The value to use when filling vectors. Only used if on_bad_vectors="fill".
         storage_options: dict, optional

python/python/lancedb/merge.py

@@ -175,9 +175,7 @@ class LanceMergeInsertBuilder(object):
             can be anything you use for [`add`][lancedb.table.Table.add]
         on_bad_vectors: str, default "error"
             What to do if any of the vectors are not the same size or contains NaNs.
-            One of "error", "drop", "fill", "null", or "keep". With "keep",
-            vectors containing NaNs are preserved, but vectors with the wrong
-            dimension still raise an error.
+            One of "error", "drop", "fill".
         fill_value: float, default 0.
             The value to use when filling vectors. Only used if on_bad_vectors="fill".
         timeout: Optional[timedelta], default None

python/python/lancedb/remote/db.py

@@ -543,9 +543,7 @@ class RemoteDBConnection(DBConnection):
             to "exist_ok".
         on_bad_vectors: str, default "error"
             What to do if any of the vectors are not the same size or contains NaNs.
-            One of "error", "drop", "fill", "null", or "keep". With "keep",
-            vectors containing NaNs are preserved, but vectors with the wrong
-            dimension still raise an error.
+            One of "error", "drop", "fill".
         fill_value: float
             The value to use when filling vectors. Only used if on_bad_vectors="fill".
 

python/python/lancedb/remote/table.py

@@ -629,9 +629,7 @@ class RemoteTable(Table):
             "append" and "overwrite".
         on_bad_vectors: str, default "error"
             What to do if any of the vectors are not the same size or contains NaNs.
-            One of "error", "drop", "fill", "null", or "keep". With "keep",
-            vectors containing NaNs are preserved, but vectors with the wrong
-            dimension still raise an error.
+            One of "error", "drop", "fill".
         fill_value: float, default 0.
             The value to use when filling vectors. Only used if on_bad_vectors="fill".
         progress: bool, callable, or tqdm-like, optional

python/python/lancedb/table.py

@@ -350,10 +350,8 @@ def _sanitize_data(
         in the input table before casting.
     metadata : Optional[dict], default None
         The embedding metadata to add to the schema.
-    on_bad_vectors : Literal["error", "drop", "fill", "null", "keep"], default "error"
+    on_bad_vectors : Literal["error", "drop", "fill", "null"], default "error"
         What to do if any of the vectors are not the same size or contains NaNs.
-        With "keep", vectors containing NaNs are preserved, but vectors with the
-        wrong dimension still raise an error.
     fill_value : float, default 0.0
         The value to use when filling vectors. Only used if on_bad_vectors="fill".
         All entries in the vector will be set to this value.
@@ -1249,9 +1247,7 @@ class Table(ABC):
             "append" and "overwrite".
         on_bad_vectors: str, default "error"
             What to do if any of the vectors are not the same size or contains NaNs.
-            One of "error", "drop", "fill", "null", or "keep". With "keep",
-            vectors containing NaNs are preserved but are not indexed for vector
-            search; vectors with the wrong dimension still raise an error.
+            One of "error", "drop", "fill".
         fill_value: float, default 0.
             The value to use when filling vectors. Only used if on_bad_vectors="fill".
         progress: bool, callable, or tqdm-like, optional
@@ -3273,9 +3269,7 @@ class LanceTable(Table):
             "append" and "overwrite".
         on_bad_vectors: str, default "error"
             What to do if any of the vectors are not the same size or contains NaNs.
-            One of "error", "drop", "fill", "null", or "keep". With "keep",
-            vectors containing NaNs are preserved, but vectors with the wrong
-            dimension still raise an error.
+            One of "error", "drop", "fill", "null".
         fill_value: float, default 0.
             The value to use when filling vectors. Only used if on_bad_vectors="fill".
         progress: bool, callable, or tqdm-like, optional
@@ -3588,9 +3582,7 @@ class LanceTable(Table):
             data but will validate against any schema that's specified.
         on_bad_vectors: str, default "error"
             What to do if any of the vectors are not the same size or contains NaNs.
-            One of "error", "drop", "fill", "null", or "keep". With "keep",
-            vectors containing NaNs are preserved, but vectors with the wrong
-            dimension still raise an error.
+            One of "error", "drop", "fill", "null".
         fill_value: float, default 0.
             The value to use when filling vectors. Only used if on_bad_vectors="fill".
         embedding_functions: list of EmbeddingFunctionModel, default None
@@ -4026,7 +4018,7 @@ class LanceTable(Table):
 
 def _handle_bad_vectors(
     reader: pa.RecordBatchReader,
-    on_bad_vectors: OnBadVectorsType = "error",
+    on_bad_vectors: Literal["error", "drop", "fill", "null"] = "error",
     fill_value: float = 0.0,
     target_schema: Optional[pa.Schema] = None,
     metadata: Optional[dict] = None,
@@ -4200,9 +4192,7 @@ def _handle_bad_vector_column(
         The name of the vector column.
     on_bad_vectors: str, default "error"
         What to do if any of the vectors are not the same size or contains NaNs.
-        One of "error", "drop", "fill", "null", or "keep". With "keep",
-        vectors containing NaNs are preserved, but vectors with the wrong dimension
-        still raise an error.
+        One of "error", "drop", "fill", "null".
     fill_value: float, default 0.0
         The value to use when filling vectors. Only used if on_bad_vectors="fill".
     """
@@ -4267,8 +4257,7 @@ def _handle_bad_vector_column(
                     f"Vector column '{vector_column_name}' has NaNs. "
                     "Set on_bad_vectors='drop' to remove them, "
                     "set on_bad_vectors='fill' and fill_value=<value> to replace them, "
-                    "set on_bad_vectors='null' to replace them with null, "
-                    "or set on_bad_vectors='keep' to preserve them."
+                    "or set on_bad_vectors='null' to replace them with null."
                 )
         elif on_bad_vectors == "null":
             vec_arr = pc.if_else(
@@ -4285,16 +4274,6 @@ def _handle_bad_vector_column(
                     "`fill_value` must not be None if `on_bad_vectors` is 'fill'"
                 )
             vec_arr = _fill_bad_vector_values(vec_arr, dim, fill_value)
-        elif on_bad_vectors == "keep":
-            if pc.any(has_wrong_dim).as_py():
-                raise ValueError(
-                    f"Vector column '{vector_column_name}' has variable length "
-                    "vectors. on_bad_vectors='keep' only preserves vectors "
-                    "containing NaNs. Set on_bad_vectors='drop' to remove "
-                    "wrong-size vectors, set on_bad_vectors='fill' and "
-                    "fill_value=<value> to replace them, or set "
-                    "on_bad_vectors='null' to replace them with null."
-                )
         else:
             raise ValueError(f"Invalid value for on_bad_vectors: {on_bad_vectors}")
 
@@ -5139,9 +5118,7 @@ class AsyncTable:
             "append" and "overwrite".
         on_bad_vectors: str, default "error"
             What to do if any of the vectors are not the same size or contains NaNs.
-            One of "error", "drop", "fill", "null", or "keep". With "keep",
-            vectors containing NaNs are preserved but are not indexed for vector
-            search; vectors with the wrong dimension still raise an error.
+            One of "error", "drop", "fill", "null".
         fill_value: float, default 0.
             The value to use when filling vectors. Only used if on_bad_vectors="fill".
         progress: callable or tqdm-like, optional
@@ -5189,7 +5166,6 @@ class AsyncTable:
                 mode or "append",
                 progress=progress,
                 write_parallelism=write_parallelism,
-                on_nan_vectors="keep" if on_bad_vectors == "keep" else None,
             )
         except RuntimeError as e:
             if "Cast error" in str(e):

python/python/lancedb/types.py

@@ -24,7 +24,7 @@ DistanceType = Literal["l2", "cosine", "dot"]
 DistanceTypeWithHamming = Literal["l2", "cosine", "dot", "hamming"]
 
 # Vector handling literals
-OnBadVectorsType = Literal["error", "drop", "fill", "null", "keep"]
+OnBadVectorsType = Literal["error", "drop", "fill", "null"]
 
 # Mode literals
 AddMode = Literal["append", "overwrite"]

python/python/tests/test_table.py

@@ -1767,40 +1767,6 @@ def test_add_with_nans(mem_db: DBConnection):
     assert np.allclose(filled_vectors[22.0], np.array([5.0, 0.0]))
 
 
-def test_add_with_non_finite_values_keep(mem_db: DBConnection):
-    schema = pa.schema([pa.field("data", pa.list_(pa.float32(), 4))])
-    table = mem_db.create_table("test", schema=schema)
-    batch = pa.table(
-        {
-            "data": pa.array(
-                [[np.nan, np.inf, -np.inf, -0.0]],
-                type=schema.field("data").type,
-            )
-        },
-        schema=schema,
-    )
-
-    with pytest.raises(ValueError, match="NaN"):
-        table.add(batch)
-
-    table.add(batch, on_bad_vectors="keep")
-
-    values = table.to_arrow()["data"][0].as_py()
-    assert np.isnan(values[0])
-    assert np.isposinf(values[1])
-    assert np.isneginf(values[2])
-    assert values[3] == 0.0
-    assert np.signbit(values[3])
-
-
-def test_add_keep_rejects_wrong_dimension(mem_db: DBConnection):
-    schema = pa.schema([pa.field("vector", pa.list_(pa.float32(), 2))])
-    table = mem_db.create_table("test", schema=schema)
-
-    with pytest.raises((ValueError, RuntimeError), match="variable length"):
-        table.add([{"vector": [1.0]}], on_bad_vectors="keep")
-
-
 def test_add_with_empty_fixed_size_list_drops_bad_rows(mem_db: DBConnection):
     class Schema(LanceModel):
         text: str

python/python/tests/test_util.py

@@ -2,7 +2,6 @@
 # SPDX-FileCopyrightText: Copyright The LanceDB Authors
 
 
-import math
 import os
 import pathlib
 from typing import Optional
@@ -365,7 +364,7 @@ def test_fill_bad_vector_values_arrow_types(vector_type, vectors, expected):
     assert actual.to_pylist() == expected
 
 
-@pytest.mark.parametrize("on_bad_vectors", ["error", "drop", "fill", "null", "keep"])
+@pytest.mark.parametrize("on_bad_vectors", ["error", "drop", "fill", "null"])
 def test_handle_bad_vectors_nan(on_bad_vectors):
     vector = pa.array([[1.0, float("nan")], [3.0, 4.0]])
     data = pa.table({"vector": vector})
@@ -380,9 +379,8 @@ def test_handle_bad_vectors_nan(on_bad_vectors):
         assert output == (
             "ValueError: Vector column 'vector' has NaNs. Set "
             "on_bad_vectors='drop' to remove them, set on_bad_vectors='fill' "
-            "and fill_value=<value> to replace them, set on_bad_vectors='null' "
-            "to replace them with null, or set on_bad_vectors='keep' to preserve "
-            "them."
+            "and fill_value=<value> to replace them, or set on_bad_vectors='null' "
+            "to replace them with null."
         )
         return
     else:
@@ -398,26 +396,10 @@ def test_handle_bad_vectors_nan(on_bad_vectors):
         expected = pa.array([[1.0, 42.0], [3.0, 4.0]])
     elif on_bad_vectors == "null":
         expected = pa.array([None, [3.0, 4.0]])
-    elif on_bad_vectors == "keep":
-        actual = output["vector"].to_pylist()
-        assert actual[0][0] == 1.0
-        assert math.isnan(actual[0][1])
-        assert actual[1] == [3.0, 4.0]
-        return
 
     assert output["vector"].combine_chunks() == expected
 
 
-def test_handle_bad_vectors_keep_rejects_wrong_dimension():
-    data = pa.table({"vector": [[1.0, 2.0], [3.0]]})
-
-    with pytest.raises(ValueError, match="only preserves vectors containing NaNs"):
-        _handle_bad_vectors(
-            data.to_reader(),
-            on_bad_vectors="keep",
-        ).read_all()
-
-
 def test_handle_bad_vectors_noop():
     # ChunkedArray should be preserved as-is
     vector = pa.chunked_array(

python/src/table.rs

@@ -21,8 +21,7 @@ use lancedb::blob::{BlobFile, BlobRangeRequest};
 use lancedb::index::scalar::FtsIndexBuilder;
 use lancedb::table::{
     AddDataMode, ColumnAlteration, Duration, FieldMetadataUpdate, FtsToken as LanceDbFtsToken,
-    NaNVectorBehavior, NewColumnTransform, OptimizeAction, OptimizeOptions, Ref,
-    Table as LanceDbTable,
+    NewColumnTransform, OptimizeAction, OptimizeOptions, Ref, Table as LanceDbTable,
 };
 use lancedb::tokenize as lancedb_tokenize;
 use pyo3::{
@@ -643,14 +642,13 @@ impl Table {
         })
     }
 
-    #[pyo3(signature = (data, mode, progress=None, write_parallelism=None, on_nan_vectors=None))]
+    #[pyo3(signature = (data, mode, progress=None, write_parallelism=None))]
     pub fn add<'a>(
         self_: PyRef<'a, Self>,
         data: PyScannable,
         mode: String,
         progress: Option<Py<PyAny>>,
         write_parallelism: Option<usize>,
-        on_nan_vectors: Option<String>,
     ) -> PyResult<Bound<'a, PyAny>> {
         let mut op = self_.inner_ref()?.add(data);
         if mode == "append" {
@@ -660,18 +658,6 @@ impl Table {
         } else {
             return Err(PyValueError::new_err(format!("Invalid mode: {}", mode)));
         }
-        match on_nan_vectors.as_deref() {
-            None | Some("error") => {}
-            Some("keep") => {
-                op = op.on_nan_vectors(NaNVectorBehavior::Keep);
-            }
-            Some(other) => {
-                return Err(PyValueError::new_err(format!(
-                    "Invalid on_nan_vectors: {}",
-                    other
-                )));
-            }
-        }
         if let Some(write_parallelism) = write_parallelism {
             op = op.write_parallelism(write_parallelism);
         }