diff --git a/milli/src/update/new/append_only_vec.rs b/milli/src/update/new/append_only_vec.rs
new file mode 100644
index 000000000..fe05dd782
--- /dev/null
+++ b/milli/src/update/new/append_only_vec.rs
@@ -0,0 +1,327 @@
+// Code taken from <https://github.com/droundy/append-only-vec/blob/main/src/lib.rs>
+// and modified in order to get a ref mut instead of the index of newly inserted items.
+
+//! AppendOnlyVec
+//!
+//! This is a pretty simple type, which is a vector that you can push into and
+//! receive a reference to the item you just inserted. The data structure never
+//! moves an element once allocated, so you can push to the vec even while holding
+//! mutable references to elements that have already been pushed.
+//!
+//! ### Scaling
+//!
+//! 1. Accessing an element is O(1), but slightly more expensive than for a
+//!    standard `Vec`.
+//!
+//! 2. Pushing a new element amortizes to O(1), but may require allocation of a
+//!    new chunk.
+//!
+//! ### Example
+//!
+//! ```
+//! use append_only_vec::AppendOnlyVec;
+//!
+//! static V: AppendOnlyVec<String> = AppendOnlyVec::<String>::new();
+//! let mut threads = Vec::new();
+//! for thread_num in 0..10 {
+//!     threads.push(std::thread::spawn(move || {
+//!          for n in 0..100 {
+//!               let s = format!("thread {} says {}", thread_num, n);
+//!               let which = V.push(s.clone());
+//!               assert_eq!(&which, &s);
+//!          }
+//!     }));
+//! }
+//!
+//! for t in threads {
+//!    t.join();
+//! }
+//!
+//! assert_eq!(V.len(), 1000);
+//! ```
+
+use std::cell::UnsafeCell;
+use std::fmt::Debug;
+use std::ptr;
+use std::sync::atomic::{AtomicUsize, Ordering};
+
+pub struct AppendOnlyVec<T> {
+    count: AtomicUsize,
+    _reserved: AtomicUsize,
+    data: [UnsafeCell<*mut T>; BITS_USED - 1 - 3],
+}
+
+unsafe impl<T: Send> Send for AppendOnlyVec<T> {}
+unsafe impl<T: Sync + Send> Sync for AppendOnlyVec<T> {}
+
+const BITS: usize = std::mem::size_of::<usize>() * 8;
+
+#[cfg(target_arch = "x86_64")]
+const BITS_USED: usize = 48;
+#[cfg(all(not(target_arch = "x86_64"), target_pointer_width = "64"))]
+const BITS_USED: usize = 64;
+#[cfg(target_pointer_width = "32")]
+const BITS_USED: usize = 32;
+
+// This takes an index into a vec, and determines which data array will hold it
+// (the first return value), and what the index will be into that data array
+// (second return value)
+//
+// The ith data array holds 1<<i values.
+const fn indices(i: usize) -> (u32, usize) {
+    let i = i + 8;
+    let bin = BITS as u32 - 1 - i.leading_zeros();
+    let bin = bin - 3;
+    let offset = i - bin_size(bin);
+    (bin, offset)
+}
+
+const fn bin_size(array: u32) -> usize {
+    (1 << 3) << array
+}
+
+#[test]
+fn test_indices() {
+    for i in 0..32 {
+        println!("{:3}: {} {}", i, indices(i).0, indices(i).1);
+    }
+    let mut array = 0;
+    let mut offset = 0;
+    let mut index = 0;
+    while index < 1000 {
+        index += 1;
+        offset += 1;
+        if offset >= bin_size(array) {
+            offset = 0;
+            array += 1;
+        }
+        assert_eq!(indices(index), (array, offset));
+    }
+}
+
+impl<T> Default for AppendOnlyVec<T> {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+impl<T> AppendOnlyVec<T> {
+    const EMPTY: UnsafeCell<*mut T> = UnsafeCell::new(ptr::null_mut());
+
+    /// Allocate a new empty array.
+    pub const fn new() -> Self {
+        AppendOnlyVec {
+            count: AtomicUsize::new(0),
+            _reserved: AtomicUsize::new(0),
+            data: [Self::EMPTY; BITS_USED - 1 - 3],
+        }
+    }
+
+    /// Find the length of the array.
+    #[inline]
+    pub fn len(&self) -> usize {
+        self.count.load(Ordering::Acquire)
+    }
+
+    fn layout(array: u32) -> std::alloc::Layout {
+        std::alloc::Layout::array::<T>(bin_size(array)).unwrap()
+    }
+
+    /// Append an element to the array and get a mutable ref to it.
+    ///
+    /// This is notable in that it doesn't require a `&mut self`, because it
+    /// does appropriate atomic synchronization.
+    pub fn push(&self, val: T) -> &mut T {
+        let idx = self._reserved.fetch_add(1, Ordering::Relaxed);
+        let (array, offset) = indices(idx);
+        let ptr = if self.len() < 1 + idx - offset {
+            // We are working on a new array, which may not have been allocated...
+            if offset == 0 {
+                // It is our job to allocate the array! The size of the array
+                // is determined in the self.layout method, which needs to be
+                // consistent with the indices function.
+                let layout = Self::layout(array);
+                let ptr = unsafe { std::alloc::alloc(layout) } as *mut T;
+                unsafe {
+                    *self.data[array as usize].get() = ptr;
+                }
+                ptr
+            } else {
+                // We need to wait for the array to be allocated.
+                while self.len() < 1 + idx - offset {
+                    std::hint::spin_loop();
+                }
+                // The Ordering::Acquire semantics of self.len() ensures that
+                // this pointer read will get the non-null pointer allocated
+                // above.
+                unsafe { *self.data[array as usize].get() }
+            }
+        } else {
+            // The Ordering::Acquire semantics of self.len() ensures that
+            // this pointer read will get the non-null pointer allocated
+            // above.
+            unsafe { *self.data[array as usize].get() }
+        };
+
+        // The contents of this offset are guaranteed to be unused (so far)
+        // because we got the idx from our fetch_add above, and ptr is
+        // guaranteed to be valid because of the loop we used above, which used
+        // self.len() which has Ordering::Acquire semantics.
+        unsafe { (ptr.add(offset)).write(val) };
+
+        // Now we need to increase the size of the vec, so it can get read. We
+        // use Release upon success, to ensure that the value which we wrote is
+        // visible to any thread that has confirmed that the count is big enough
+        // to read that element. In case of failure, we can be relaxed, since
+        // we don't do anything with the result other than try again.
+        while self
+            .count
+            .compare_exchange(idx, idx + 1, Ordering::Release, Ordering::Relaxed)
+            .is_err()
+        {
+            // This means that someone else *started* pushing before we started,
+            // but hasn't yet finished. We have to wait for them to finish
+            // pushing before we can update the count. Note that using a
+            // spinloop here isn't really ideal, but except when allocating a
+            // new array, the window between reserving space and using it is
+            // pretty small, so contention will hopefully be rare, and having a
+            // context switch during that interval will hopefully be vanishingly
+            // unlikely.
+            std::hint::spin_loop();
+        }
+
+        unsafe { &mut *ptr }
+    }
+
+    /// Convert into a standard `Vec`.
+    pub fn into_vec(self) -> Vec<T> {
+        let mut vec = Vec::with_capacity(self.len());
+
+        for idx in 0..self.len() {
+            let (array, offset) = indices(idx);
+            // We use a Relaxed load of the pointer, because the loop above (which
+            // ends before `self.len()`) should ensure that the data we want is
+            // already visible, since it Acquired `self.count` which synchronizes
+            // with the write in `self.push`.
+            let ptr = unsafe { *self.data[array as usize].get() };
+
+            // Copy the element value. The copy remaining in the array must not
+            // be used again (i.e. make sure we do not drop it)
+            let value = unsafe { ptr.add(offset).read() };
+
+            vec.push(value);
+        }
+
+        // Prevent dropping the copied-out values by marking the count as 0 before
+        // our own drop is run
+        self.count.store(0, Ordering::Relaxed);
+
+        vec
+    }
+}
+
+impl<T> Debug for AppendOnlyVec<T> {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("AppendOnlyVec").field("len", &self.len()).finish()
+    }
+}
+
+impl<T> Drop for AppendOnlyVec<T> {
+    fn drop(&mut self) {
+        // First we'll drop all the `T` in a slightly sloppy way.  FIXME this
+        // could be optimized to avoid reloading the `ptr`.
+        for idx in 0..self.len() {
+            let (array, offset) = indices(idx);
+            // We use a Relaxed load of the pointer, because the loop above (which
+            // ends before `self.len()`) should ensure that the data we want is
+            // already visible, since it Acquired `self.count` which synchronizes
+            // with the write in `self.push`.
+            let ptr = unsafe { *self.data[array as usize].get() };
+            unsafe {
+                ptr::drop_in_place(ptr.add(offset));
+            }
+        }
+        // Now we will free all the arrays.
+        for array in 0..self.data.len() as u32 {
+            // This load is relaxed because no other thread can have a reference
+            // to Self because we have a &mut self.
+            let ptr = unsafe { *self.data[array as usize].get() };
+            if !ptr.is_null() {
+                let layout = Self::layout(array);
+                unsafe { std::alloc::dealloc(ptr as *mut u8, layout) };
+            } else {
+                break;
+            }
+        }
+    }
+}
+
+impl<T> IntoIterator for AppendOnlyVec<T> {
+    type Item = T;
+    type IntoIter = std::vec::IntoIter<T>;
+
+    fn into_iter(self) -> Self::IntoIter {
+        self.into_vec().into_iter()
+    }
+}
+
+#[test]
+fn test_parallel_pushing() {
+    use std::sync::Arc;
+    let v = Arc::new(AppendOnlyVec::<u64>::new());
+    let mut threads = Vec::new();
+    const N: u64 = 100;
+    for thread_num in 0..N {
+        let v = v.clone();
+        threads.push(std::thread::spawn(move || {
+            let which1 = v.push(thread_num);
+            let which2 = v.push(thread_num);
+            assert_eq!(*which1, thread_num);
+            assert_eq!(*which2, thread_num);
+        }));
+    }
+    for t in threads {
+        t.join().unwrap();
+    }
+    let v = Arc::into_inner(v).unwrap().into_vec();
+    for thread_num in 0..N {
+        assert_eq!(2, v.iter().copied().filter(|&x| x == thread_num).count());
+    }
+}
+
+#[test]
+fn test_into_vec() {
+    struct SafeToDrop(bool);
+
+    impl Drop for SafeToDrop {
+        fn drop(&mut self) {
+            assert!(self.0);
+        }
+    }
+
+    let v = AppendOnlyVec::new();
+
+    for _ in 0..50 {
+        v.push(SafeToDrop(false));
+    }
+
+    let mut v = v.into_vec();
+    assert_eq!(v.len(), 50);
+
+    for i in v.iter_mut() {
+        i.0 = true;
+    }
+}
+
+#[test]
+fn test_push_then_index_mut() {
+    let v = AppendOnlyVec::<usize>::new();
+    for i in 0..1024 {
+        *v.push(i) += 1;
+    }
+
+    let v = v.into_vec();
+    for i in 0..1024 {
+        assert_eq!(v[i], 2 * i);
+    }
+}
diff --git a/milli/src/update/new/mod.rs b/milli/src/update/new/mod.rs
index 98b60378f..6a48e0407 100644
--- a/milli/src/update/new/mod.rs
+++ b/milli/src/update/new/mod.rs
@@ -5,6 +5,7 @@ pub use top_level_map::{CowStr, TopLevelMap};
 use super::del_add::DelAdd;
 use crate::FieldId;
 
+mod append_only_vec;
 mod channel;
 mod document_change;
 mod extract;