diff --git a/docs/rde-import-architecture.md b/docs/rde-import-architecture.md
new file mode 100644
index 000000000..66aa57828
--- /dev/null
+++ b/docs/rde-import-architecture.md
@@ -0,0 +1,74 @@
+# Registry Data Import Architecture
+
+*See also the [RDE usage guide](./rde-import-usage.md).*
+
+The Registry Data Import feature was designed to handle escrow files from other
+registries with millions of domains. In the spirit of divide and conquer, the
+mapreduce library is used to break up the work of the import into smaller chunks
+that can be processed in a reasonable period of time. This process is broken
+down into four separate mapreduce jobs that must be run in sequence due to how
+datastore transactions work and due to dependencies between registry objects.
+The steps are broken up as follows:
+
+__Initial Setup__ - This is a set of manual steps that must be completed before
+the process can be run, and is out of the scope of this document. See
+[Usage](./rde-import-usage.md) for more details.
+
+__Contacts Import__ - Reads contact entries from an escrow file and saves them
+as `ContactResource` entities. `HistoryEntry` entities are also created for the
+contact. This process depends on initial setup, but does not depend on any
+previous step being run.
+
+__Hosts Import__ - Reads host entries from an escrow file and saves them as
+`HostResource` entities. `HistoryEntry` entities are also created for the hosts.
+This process depends on initial setup, but does not depend on any previous step
+being run.
+
+__Domains Import__ - Reads domain entries from an escrow file and saves them as
+`DomainResource` entities. For each domain imported, a history entry, autorenew
+billing event and autorenew poll message will also be created. For domains that
+are in pending transfer state, the import process will also create future
+entities for automatic server approval in the same fashion as domain transfer
+request EPP messages. Domains cannot be imported until the contacts and hosts
+required by the domain are imported in previous steps.
+
+__Hosts Link__ - Reads host entries from an escrow file and links in-zone hosts
+to their superordinate domains. This is the last step because both hosts and
+domains have to be imported before the link can be made in both directions.
+
+## Components
+
+Each mapreduce job (with the exception of Hosts Link) is made up of a similar
+set of components. Note that much of the work that is done by each job strongly
+resembles the inversion of the Registry Data Export feature, and reuses the Jaxb
+representations of the xml elements that compose escrow files.
+
+__Parser__ - The parser is the lowest level of the import process. This
+component parses an escrow file (provided as an open stream from Google Cloud
+Storage) into discrete JAXB objects. The parser maintains an internal cursor in
+the xml file that represents the next element to be read, and can advance to and
+skip any number of elements.
+
+__Reader__ - The reader is configured by each mapreduce job to load an escrow
+file from Cloud Storage and use a parser to read a selected subset of the file,
+forwarding the results to a mapper.
+
+__Input__ - An input is responsible for determining how many reader instances to
+create and which section of the escrow file should be consumed by each reader.
+
+__Converter__ - The converter accepts a Jaxb object and returns an equivalent
+resource that can be saved to the datastore.
+
+__Import Utility Logic__ - Common import logic is consolidated into a single
+place, such as creation of index entities and escrow file validation.
+
+__Mapper__ - The mapper accepts a stream of Jaxb objects from the reader and
+uses the converter to map them to resource objects. Then for each resource
+object produced, the mapper will attempt to save the resource and any related
+objects to the datastore in a transaction. This is an idempotent operation; if
+any resource has been previously imported by the process, it will be ignored.
+
+__Action Endpoint__ - The action endpoint is responsible for accepting requests
+to launch each step of the import process, bootstrapping mapreduce jobs, and
+redirecting the client to the status page of the import job. This is the entry
+point of the import process.