Note some subtasks and difficulties involved with reducing RAM usage on dirservers

svn:r6428

1 files changed, 21 insertions, 2 deletions

diff --git a/doc/TODO b/doc/TODO
index 24c322cbf6..496dc1214f 100644
--- a/doc/TODO
+++ b/doc/TODO
@@ -62,11 +62,30 @@ Must-have items for 0.1.2.x:
 R - Server usability
 N - Better hidden service performance, with possible redesign.
   - Asynchronous DNS
+    - What to use? C-ares? Libdns? AGL's patch?
   - Better estimates in the directory of whether servers have good uptime
     (high expected time to failure) or good guard qualities (high
     fractional uptime).
-  - memory usage on dir servers.
-    copy less!
+N . memory usage on dir servers. copy less!
+    o Remember offset and location of each descriptor in the cache/journal
+    - When sending a big pile of descs to a client, don't shove them all on
+      the buffer at once.
+      X This may require routerinfo_t or signed_descriptor_t to get slightly
+        refcounted.  (Only slightly; we'd only need to know whether it's on
+        the routerlist->routers or routerlist->old_routers, and how many
+        directory conns are returning it.)
+      - Possibly, we could just to keep a list of the descriptor digests for
+        the descriptors we still want to send.  We might end up truncating
+        some replies by returning fewer descriptors than were requested (if
+        somebody requests a desc that we throw away before we deliver it),
+        but this happens only when somebody wants an obsolete desc, and
+        clients can already handle truncated replies.
+      - But what do we do about compression? That's the part that makes
+        stuff hard.
+
+    - Make sure offset is correct in the presence of windows FS insanity.
+    - Consider whether it's smart to mmap cache files where possible.
+    - Consider whether it's smart to lazy-load routerdescs on non-directories.
 N - oprofile including kernel time on multiple platforms
 
 Items for 0.1.2: