diff --git a/apps/emqx_durable_storage/README.md b/apps/emqx_durable_storage/README.md
index bc8eae2d0..f613085bb 100644
--- a/apps/emqx_durable_storage/README.md
+++ b/apps/emqx_durable_storage/README.md
@@ -1,24 +1,97 @@
-# EMQX Replay
+# EMQX Durable Storage
 
-`emqx_ds` is an application implementing durable storage for MQTT messages within EMQX.
+`emqx_durable_storage` (DS for short) is an application implementing durable storage for MQTT messages within EMQX.
+
+The core design idea behind `emqx_durable_storage` is to store each message exactly once (per each replica of the database), regardless of the number of consumers, online or offline.
+This makes the storage disk requirements very predictable: only the number of _published_ messages matters; the number of consumers is removed from the equation, and fan-out is practically free in terms of disk storage.
 
 # Features
 
-- Streams. Stream is an abstraction that encompasses topics, shards, different data layouts, etc.
-  The client application must only aware of the streams.
+## Callback modules
 
-- Batching. All the API functions are batch-oriented.
+### Backend
 
-- Iterators. Iterators can be stored durably or transferred over network.
-  They take relatively small space.
+DS _backend_ is a callback module that implements `emqx_ds` behavior.
 
-- Support for various backends. Almost any DBMS that supports range
-  queries can serve as a `emqx_durable_storage` backend.
+EMQX repository contains the "builtin" backend, implemented in `emqx_ds_replication_layer` module, that uses RocksDB as the main storage.
 
-- Builtin backend based on RocksDB.
-  - Changing storage layout on the fly: it's achieved by creating a
-    new set of tables (known as "generation") and the schema.
-  - Sharding based on publisher's client ID
+Note that builtin backend introduces the concept of **site** to alleviate the problem of changing node names.
+Site IDs are persistent, and they are randomly generated at the first startup of the node.
+Each node in the cluster has a unique site ID, that is independent from the Erlang node name (`emqx@...`).
+
+### Layout
+
+DS _layout_ is a module that implements `emqx_ds_storage_layer` behavior.
+Layout modules are the only modules that have direct access to the underlying storage engine, in both reads and writes.
+
+Different storage layouts can be used to maximize the efficiency of message storage and retrieval for various types of workload.
+
+Backward- and forward-incompatible changes to the layout modules are forbidden.
+EMQX should always be able to read the data written by the old releases.
+Non-compatible changes must be implemented as entirely new layout modules.
+
+## How does EMQX organize data
+
+Messages are organized in the following hierarchy:
+
+1. **Database**.
+   DS databases are completely independent from each other.
+   They can have different schema, different backend, and they can be opened, closed, created and dropped independently from each other.
+
+   Each database can be used for a different type of payload or a different tenant.
+
+2. **Shard**.
+   (The concept of shard is specific to the builtin backend)
+   The builtin backend separates different messages into shards.
+   Sharding can be performed by `clientId` or the topic of the message.
+
+3. **Generation**.
+   Each shard is additionally split into partitions called generations, each one covering a particular period of time.
+   New messages are written only into the _current_ generation, while the previous generations are only accessible for reading.
+
+   Different generations can use different layout modules to organize the data.
+   In fact, in order to change the layout of the data the application must create a new generation, so the previously recorded messages remain readable without having to perform a heavy migration procedure.
+   Generations can also be used for the garbage collection and message retention policies: since all messages in the generation belong to a certain interval of time, old messages can be efficiently deleted by dropping the entire generation.
+
+
+4. *Stream*.
+   Finally, messages in each shard and generation are split into streams.
+   Every stream can contain messages from multiple topics.
+   The number of streams is expected to be relatively low in comparison with the number of topics: one stream can potentially contain millions of topics.
+   Various layout callback modules can employ different strategies for mapping topics into streams.
+
+   Stream is *the only* unit of message serialization in `emqx_durable_storage` application.
+
+   The consumer of the messages can replay the stream using an _iterator_.
+
+## Message replay
+
+All the API functions in EMQX DS are batch-oriented.
+
+Consumption of messages is done in several stages:
+
+1. The consumer calls `emqx_ds:get_streams` function to get the list of streams that contain messages from a given topic filter, and a given time range.
+
+2. `get_streams` returns the list of streams together with their _ranks_.
+   The rank of the stream is a tuple with two elements, called `X` and `Y`.
+
+   The consumer must follow the below rules to avoid reordering of the messages:
+
+   - Streams with different `X`-ranks can always be replayed in parallel, regardless of their `Y`-rank.
+   - Streams with the same `X` and `Y`-rank can be replayed in parallel.
+   - Groups of streams with the same `X` rank should be replayed in order of their `Y`-rank
+
+3. In order to start replay of the stream, the consumer calls `emqx_ds:make_iterator` function that returns an _iterator_ object.
+   Iterators are the pointers to a particular position in the stream, they can be saved and restored as regular Erlang terms.
+
+4. The consumer then proceeds to call `emqx_ds:next` function to fetch messages.
+   - If this function returns `{ok, end_of_stream}`, it means the stream is fully replayed.
+   - If this function returns `{ok, NextIterator, []}`, it means new messages can appear in the stream.
+
+   Note: the consumer must implement a fair strategy for consuming messages from different streams.
+   It cannot rely on an assumption that it can reach the end of a stream in a finite time.
+
+5. The consumer must periodically refresh the list of streams as explained in 1, because new streams can appear from time to time.
 
 # Limitation
 
@@ -36,10 +109,16 @@ In the future it can serve as a storage for retained messages or as a generic me
 
 # Configurations
 
-`emqx_durable_storage` doesn't have any configurable parameters.
-Instead, it relies on the upper-level business applications to create
-a correct configuration and pass it to `emqx_ds:open_db(DBName, Config)`
-function according to its needs.
+Global options for `emqx_durable_storage` application are configured via OTP application environment.
+Database-specific settings are stored in the schema table.
+
+The following application environment variables are available:
+
+- `emqx_durable_storage.db_data_dir`: directory where the databases are located
+
+- `emqx_durable_storage.egress_batch_size`: number of messages stored in the batch before it is committed to the durable storage.
+
+- `emqx_durable_storage.egress_flush_interval`: period at which the batches of messages are committed to the durable storage.
 
 # HTTP APIs