diff --git a/docs/production/requirements.md b/docs/production/requirements.md
index 335f1ea450..6d9093e863 100644
--- a/docs/production/requirements.md
+++ b/docs/production/requirements.md
@@ -230,9 +230,10 @@ installing Zulip with a dedicated database server.
   backend][s3-uploads].
 
 - **Sharding:** For servers with several thousand daily active users,
-  Zulip supports sharding its real-time-push Tornado service, both by
-  realm/organization (for hosting many organizations) and by user ID
-  (for hosting single very large organizations)
+  Zulip supports [sharding its real-time-push Tornado
+  service][tornado-sharding], both by realm/organization (for hosting many
+  organizations) and by user ID (for hosting single very large
+  organizations).
 
   Care must be taken when dividing traffic for a single Zulip realm
   between multiple Zulip application servers, which is why we
@@ -250,3 +251,4 @@ document](../subsystems/performance.md) may also be of interest.
 [s3-uploads]: upload-backends.md#s3-backend-configuration
 [streaming-replication]: postgresql.md#postgresql-warm-standby
 [contact-support]: https://zulip.com/help/contact-support
+[tornado-sharding]: system-configuration.md#tornado_sharding
diff --git a/docs/production/system-configuration.md b/docs/production/system-configuration.md
index bd16e30727..0d0b51da40 100644
--- a/docs/production/system-configuration.md
+++ b/docs/production/system-configuration.md
@@ -361,6 +361,48 @@ Set to a true value to enable object size reporting in memcached. This incurs a
 small overhead for every store or delete operation, but allows a
 memcached_exporter to report precise item size distribution.
 
+### `[tornado_sharding]`
+
+Keys in this section are used to configure how many Tornado instances
+are started, and which users are mapped to which of those instances.
+Each Tornado instance listens on a separate port, starting at 9800 and
+proceeding upwards from there. A single Tornado instance can usually
+handle 1000-1500 concurrent active users, depending on message sending
+volume.
+
+Individual organizations may be assigned to ports, either via their
+subdomain names, or their fully-qualified hostname (for [organizations
+using `REALM_HOSTS`][multiple-organizations.md#other-hostnames]):
+
+```ini
+[tornado_sharding]
+9800 = realm-a realm-b
+9801 = realm-c
+9802 = realm-host.example.net
+```
+
+Organizations can also be assigned to ports via regex over their
+fully-qualified hostname:
+
+```ini
+[tornado_sharding]
+9800_regex = ^realm-(a|b)\.example\.com$
+9801_regex = ^other(-option)?\.example.com$
+```
+
+Extremely large organizations can be distributed across multiple
+Tornado shards by joining the ports in the key with `_`:
+
+```ini
+[tornado_sharding]
+9800 = small-realm
+9801_9802 = very-large-realm
+```
+
+After running `scripts/zulip-puppet-apply`, a separate step to run
+`scripts/refresh-sharding-and-restart` is required for any sharding
+changes to take effect.
+
 ### `[loadbalancer]`
 
 #### `ips`
diff --git a/docs/subsystems/performance.md b/docs/subsystems/performance.md
index 016e532d41..72e9d333eb 100644
--- a/docs/subsystems/performance.md
+++ b/docs/subsystems/performance.md
@@ -137,12 +137,11 @@ idle for a minute. It's likely that with some strategy for detecting
 such situations, we could reduce their volume (and thus overall
 Tornado load) dramatically.
 
-Currently, Tornado is sharded by realm, which is sufficient for
-arbitrary scaling of the number of organizations on a multi-tenant
-system like zulip.com. With a somewhat straightforward set of work,
-one could change this to sharding by `user_id` instead, which will
-eventually be important for individual large organizations with many
-thousands of concurrent users.
+Currently, Tornado is sharded by realm, and optionally by user-id
+within each realm. Sharding by realm is sufficient for arbitrary
+scaling of the number of organizations on a multi-tenant system like
+zulip.com. Sharding by user-id is necessary for very large
+organizations with multiple thousands of active users at once.
 
 ### Presence