diff --git a/design.md b/design.md
index 580db36..50fe6d1 100644
--- a/design.md
+++ b/design.md
@@ -453,3 +453,36 @@ impl Writer for IoUringLocal { }
 impl GetFilesize for IoUringLocal { }
 impl ReadThenWrite for IoUringLocal { }
 ```
+
+### Trying out some iterator ideas again
+
+```rust
+let errors = reader.read(
+    vec![
+        // First group: All the source GRIB files which will be collated into
+        // a set of Zarr chunks.
+        HashMap::from([
+            ("/foo/grib1.1", ByteRanges(vec![...])),
+            ("/foo/grib1.2", ByteRanges(vec![...])),
+        ]),
+        // Second group:
+        HashMap::from([
+            ("/foo/grib2.1", ByteRanges(vec![...])),
+            ("/foo/grib2.1", ByteRanges(vec![...])),
+        ]),
+    ])
+    // Behind the scenes, `reader.read()` launches a thread which owns its own
+    // io_uring, and continually keeps that io_uring topped up. All IO operations in group n
+    // are guaranteed to be completed before LSIO begins IO ops from group n+1.
+    .iter()
+    // `map` acts on each byte range.
+    // TODO: How to allow `map` to run concurrently across all tasks? Maybe:
+    //    - Rayon's par_iter?
+    //    - Return a Future? (Not sure that'll work?)
+    //    - Use Tokio with Rayon?! See https://ryhl.io/blog/async-what-is-blocking/
+    .map(|(buffer, path, byte_range, group_index)| (decompress(buffer), path, byte_range, group_index))
+    // `reduce` collects all the buffers for a group, and outputs a vector of
+    // (output_buffer, output_path, output_byte_range)
+    .reduce(|buffers_with_paths_and_byte_range| collate(buffers_with_paths_and_byte_range))
+    // Now compress each 
+```