diff --git a/docs/source/mldatasets.md b/docs/source/mldatasets.md
new file mode 100644
index 000000000..360153727
--- /dev/null
+++ b/docs/source/mldatasets.md
@@ -0,0 +1,87 @@
+xcube Multi-Resolution Datasets
+===============================
+
+Definition
+----------
+
+A xcube _multi-resolution dataset_ refers to an N-D [image 
+pyramid](https://en.wikipedia.org/wiki/Pyramid_(image_processing)) 
+where an _image_ refers to a 2-D dataset with two spatial dimensions
+in some horizontal coordinate system.
+
+A multi-resolution dataset comprises a fixed number of
+_levels_, which are regular datasets covering the same spatial area at different resolutions.
+Level zero represents the original resolution `res(L=0)`, higher level 
+resolutions decrease by a factor of two: `res(L) = res(0) / 2^L`.
+
+
+Implementation in xcube
+-----------------------
+
+In xcube, multi-resolution datasets are represented by the abstract class
+`xcube.core.mldataset.MultiLevelDataset`. The xcube data store framework
+refers to this datatype using the alias `mldataset`. The corresponding
+default data format is the xcube `levels` format. Later xcube will also 
+support Cloud Optimized GeoTIFF (COG) as format for multi-resolution 
+datasets.
+
+The xcube Levels Format
+-----------------------
+
+The xcube Levels format is basically a single top-level directory. 
+The filename extension of that directory should be `.levels` 
+by convention. The directory entries are Zarr datasets   
+
+1. that are representations of regular xarray datasets named after 
+   their zero-based level index, `{level}.zarr`;
+2. that comply with the xcube dataset convention.
+
+<div style="color: red;">
+    TODO (forman): link to xcube dataset convention
+</div> 
+
+
+The following is a multi-resolution dataset with three levels:
+
+    - test_pyramid.levels/
+        - 0.zarr/
+        - 1.zarr/
+        - 2.zarr/
+
+An important use case is generating image pyramids from existing large 
+datasets without the need to create a copy of level zero.
+
+To support this, the level zero dataset may be a link to an existing 
+Zarr dataset. The filename is then `0.link` rather than `0.zarr`. 
+The link file contains the path to the actual Zarr dataset 
+to be used as level zero as a plain text string. It may be an absolute 
+path or a path relative to the top-level dataset.
+
+    - test_pyramid.levels/
+        - 0.link/    # --> link to actual level zero dataset
+        - 1.zarr/
+        - 2.zarr/
+
+Related reads
+-------------
+
+* [WIP: Multiscale use-case](https://github.com/zarr-developers/zarr-specs/issues/23)
+  in zarr-developers / zarr-specs on GitHub
+
+
+To be discussed
+---------------
+
+* Allow links for all levels?
+* Add top-level metadata such as `num_levels` and links for each 
+  level?
+* Make top-level directory a Zarr group (`.zgroup`)
+  and encode level metadata in `.zattrs` (e.g. `num_levels`)?
+* Link relative to link file?
+
+To do
+-----
+
+* Currently, the FS data stores treat relative link paths as relative
+  to the data store's `root`. 
+