diff --git a/docs/restate/architecture.md b/docs/restate/architecture.md
index 784dd0e4..2bb080fc 100644
--- a/docs/restate/architecture.md
+++ b/docs/restate/architecture.md
@@ -11,7 +11,7 @@ The runtime is split into the control plane consisting of *Metas* and the data p
 
 The *Metas* are responsible for managing the service meta information and coordinating the *Workers*.
 
-The *Workers* are responsible for invoking services, storing their journal and service state as well as maintaining processing order.
+The *Workers* are responsible for invoking services, storing the invocation and service state as well as maintaining processing order.
 *Workers* also expose a *SQL* interface to query the runtime's internal state.
 
 Services run on *Service Endpoints* which can be deployed independently of the runtime.
@@ -56,3 +56,48 @@ Moreover, it supports asynchronous checkpointing which is required to truncate t
 Restate makes its internal state accessible via a SQL interface.
 Any client that supports the PostgreSQL wire protocol can be used to issue queries.
 Internally, the SQL queries are executed using [DataFusion](https://github.com/apache/arrow-datafusion).
+
+## Durable execution via journaling
+
+Things always eventually go wrong, in particular at scale when running a distributed application.
+Therefore, services that Restate invokes will also fail eventually.
+In these cases, the services need to be re-invoked so that they can complete.
+Ideally, services resume from the point where they failed without re-doing all the previous actions, effectively resulting into durable execution of services.
+
+The way Restate achieves durable execution is by recording the actions a service takes in a durable journal.
+When re-invoking a service, the journal is sent alongside the service state to the service endpoint.
+The service endpoint resumes the invocation by replaying the recorded journal entries without re-doing the actual operations (e.g. calling another service, storing state, running a side-effect, etc.).
+A useful side-effect of the journal is that Restate can always suspend an invocation and resume it at a later point in time if need should arise.
+
+![Durable execution](/img/durable-execution.png)
+
+## Service registry
+
+All servie meta information is maintained by the *Metas* via the service registry.
+The service registry contains information about the registered services which includes the address of the service endpoints, the exposed service methods, their signatures and type definitions.
+The service contracts and their type definitions are also exposed via [gRPC reflection](https://github.com/grpc/grpc/blob/master/doc/server-reflection.md) by the *Workers*.
+
+Services are added to the registry by discovering service endpoints.
+Upon discovery request, the *Metas* reach out to the specified service endpoint and retrieve all registered services, their methods and type definitions.
+After discovery, the new service meta information is propagated to the *Workers* which enables invoking these services through Restate.
+
+## Service invocation flow
+
+All service invocations need to go through Restate which acts as a reverse-proxy for the registered services.
+Restate knows where the services are running and invokes them on behalf of the clients.
+By putting Restate in charge of the actual invocation, the system can achieve the following properties:
+
+* Enrich the invocation with extra information such as service and journal state
+* Guarantee that keyed invocations are executed in order and isolation
+* Retry failed invocations until they complete
+
+The service invocation flow is as follows:
+
+1. New invocations are received by the *Ingress service* running on the *Workers*. The ingress extracts the invocation key which determines the partition responsible for processing the invocation.
+2. Based on the key, the invocation is forwarded to the right *Worker* running the *Partition processor* of the target partition.
+3. The *Partition processor* makes sure that for a given key and service, the invocations are executed sequentially.
+4. Once all preceding invocations have completed, the invocation request will be enriched with the service's state and sent to the service endpoint where it executes.
+5. While executing, the runtime journals actions the service takes in order to be able to recover from failures without redoing these actions.
+6. Once the invocation completes with a response, it will be sent back to the *Ingress service* and then to the invoking client.
+
+![Service invocation flow](/img/service-invocation-flow.png)
diff --git a/static/img/durable-execution.png b/static/img/durable-execution.png
new file mode 100644
index 00000000..eae6864e
Binary files /dev/null and b/static/img/durable-execution.png differ
diff --git a/static/img/service-invocation-flow.png b/static/img/service-invocation-flow.png
new file mode 100644
index 00000000..2089c0e0
Binary files /dev/null and b/static/img/service-invocation-flow.png differ