deploy: cde8af9a495cbc7f3d0207e3f17c37eddaee34e1

2 files changed, 37 insertions, 1 deletions

diff --git a/develop/usage/configuration/config_documentation.html b/develop/usage/configuration/config_documentation.html
index ab92ac22ce..10bcd93e9a 100644
--- a/develop/usage/configuration/config_documentation.html
+++ b/develop/usage/configuration/config_documentation.html
@@ -1049,14 +1049,32 @@ with intermittent connections, at the cost of higher memory usage.
 By default, this is zero, which means that sync responses are not cached
 at all.</p>
 </li>
+<li>
+<p><code>cache_autotuning</code> and its sub-options <code>max_cache_memory_usage</code>, <code>target_cache_memory_usage</code>, and
+<code>min_cache_ttl</code> work in conjunction with each other to maintain a balance between cache memory 
+usage and cache entry availability. You must be using <a href="https://github.com/matrix-org/synapse#help-synapse-is-slow-and-eats-all-my-ramcpu">jemalloc</a> 
+to utilize this option, and all three of the options must be specified for this feature to work.</p>
+<ul>
+<li><code>max_cache_memory_usage</code> sets a ceiling on how much memory the cache can use before caches begin to be continuously evicted.
+They will continue to be evicted until the memory usage drops below the <code>target_memory_usage</code>, set in
+the flag below, or until the <code>min_cache_ttl</code> is hit.</li>
+<li><code>target_memory_usage</code> sets a rough target for the desired memory usage of the caches.</li>
+<li><code>min_cache_ttl</code> sets a limit under which newer cache entries are not evicted and is only applied when
+caches are actively being evicted/<code>max_cache_memory_usage</code> has been exceeded. This is to protect hot caches
+from being emptied while Synapse is evicting due to memory.</li>
+</ul>
+</li>
 </ul>
 <p>Example configuration:</p>
 <pre><code class="language-yaml">caches:
   global_factor: 1.0
   per_cache_factors:
     get_users_who_share_room_with_user: 2.0
-  expire_caches: false
   sync_response_cache_duration: 2m
+  cache_autotuning:
+    max_cache_memory_usage: 1024M
+    target_cache_memory_usage: 758M
+    min_cache_ttl: 5m
 </code></pre>
 <h3 id="reloading-cache-factors"><a class="header" href="#reloading-cache-factors">Reloading cache factors</a></h3>
 <p>The cache factors (i.e. <code>caches.global_factor</code> and <code>caches.per_cache_factors</code>)  may be reloaded at any time by sending a
diff --git a/develop/usage/configuration/homeserver_sample_config.html b/develop/usage/configuration/homeserver_sample_config.html
index b412f1caa3..c951cb8dc2 100644
--- a/develop/usage/configuration/homeserver_sample_config.html
+++ b/develop/usage/configuration/homeserver_sample_config.html
@@ -940,6 +940,24 @@ caches:
   #
   #cache_entry_ttl: 30m
 
+  # This flag enables cache autotuning, and is further specified by the sub-options `max_cache_memory_usage`,
+  # `target_cache_memory_usage`, `min_cache_ttl`. These flags work in conjunction with each other to maintain
+  # a balance between cache memory usage and cache entry availability. You must be using jemalloc to utilize
+  # this option, and all three of the options must be specified for this feature to work.
+  #cache_autotuning:
+    # This flag sets a ceiling on much memory the cache can use before caches begin to be continuously evicted.
+    # They will continue to be evicted until the memory usage drops below the `target_memory_usage`, set in
+    # the flag below, or until the `min_cache_ttl` is hit.
+    #max_cache_memory_usage: 1024M
+
+    # This flag sets a rough target for the desired memory usage of the caches.
+    #target_cache_memory_usage: 758M
+
+    # 'min_cache_ttl` sets a limit under which newer cache entries are not evicted and is only applied when
+    # caches are actively being evicted/`max_cache_memory_usage` has been exceeded. This is to protect hot caches
+    # from being emptied while Synapse is evicting due to memory.
+    #min_cache_ttl: 5m
+
   # Controls how long the results of a /sync request are cached for after
   # a successful response is returned. A higher duration can help clients with
   # intermittent connections, at the cost of higher memory usage.