diff --git a/DESIGN-connections.md b/DESIGN-connections.md
new file mode 100644
index 0000000..d8e5eaf
--- /dev/null
+++ b/DESIGN-connections.md
@@ -0,0 +1,76 @@
+Connection closing and subsequent re-opening is not free. I've done some experiments; it looks like the connection
+close/open themselves are in the sub-1ms range.
+
+But: when being the first connection to write, there is a c. 13ms penalty. Hypothesis: the opening of the WAL file.
+
+This is the code I played with to establish this.
+Note that the result of running the second half (the Django part) without first running the first half are very
+different from doing them both at once. Hence "being the first connection", i.e. there's a cross-connection effect.
+
+```
+from performance.context_managers import time_it
+import sqlite3
+
+conn = sqlite3.connect('db.sqlite3')
+cursor = conn.cursor()
+
+with time_it() as timings:
+  cursor.execute('UPDATE auth_user set username = username')
+  conn.commit()
+
+print("fresh conn", timings.took)
+
+with time_it() as timings:
+  cursor.execute('UPDATE auth_user set username = username')
+  conn.commit()
+
+print("reused conn", timings.took)
+
+
+from performance.context_managers import time_it
+import sqlite3
+from django.db import connection
+from django.db import models
+from users.models import User
+from performance.context_managers import time_it
+from bugsink.transaction import immediate_atomic
+
+with time_it() as timings:
+    with immediate_atomic():
+        User.objects.update(username=models.F('username'))
+
+print("django fresh conn", timings.took)
+
+with time_it() as timings:
+    with immediate_atomic():
+        User.objects.update(username=models.F('username'))
+
+print("django reused conn", timings.took)
+```
+
+With just the second half:
+
+```
+...     with immediate_atomic():
+...         User.objects.update(username=models.F('username'))
+... 
+CONNECTION created /mnt/datacrypt/dev/bugsink/db.sqlite3 124812146634816 in 0ms MainThread
+      1.79ms BEGIN IMMEDIATE, A.K.A. get-write-lock ↴
+      0.69ms ABOUT TO END IMMEDIATE transaction' ↴
+     10.67ms END IMMEDIATE transaction' ↴
+>>> print("django fresh conn", timings.took)
+
+django fresh conn 13.312101364135742
+
+>>> 
+>>> with time_it() as timings:
+...     with immediate_atomic():
+...         User.objects.update(username=models.F('username'))
+... 
+      0.19ms BEGIN IMMEDIATE, A.K.A. get-write-lock ↴
+      0.35ms ABOUT TO END IMMEDIATE transaction' ↴
+      0.50ms END IMMEDIATE transaction' ↴
+>>> print("django reused conn", timings.took)
+
+django reused conn 0.9033679962158203
+```
diff --git a/snappea/foreman.py b/snappea/foreman.py
index e9a98f8..cd99e4b 100644
--- a/snappea/foreman.py
+++ b/snappea/foreman.py
@@ -150,6 +150,8 @@ class Foreman:
         # those constructs stands out. And outside of such decorators you're in autocommit because that's what Django
         # does.
         #
+        # Connection close (and subsequent reopen) has a cost that's documented in DESIGN-connections.md.
+        #
         # Calling connection.close unnecessarily does not incur extra cost, because of `if self.connection is not None`
         # in django.db.backends.base.base.DatabaseWrapper. (this is also the reason repeated calls to connection.close()
         # are not a problem). (`connection` below is actually a ConnectionProxy, but it delegates to a DatabaseWrapper).