diff --git a/atlasdb-cassandra/src/main/java/com/palantir/atlasdb/cassandra/CassandraKeyValueServiceConfig.java b/atlasdb-cassandra/src/main/java/com/palantir/atlasdb/cassandra/CassandraKeyValueServiceConfig.java
index e9aad7eeee1..f29e11ef4dc 100644
--- a/atlasdb-cassandra/src/main/java/com/palantir/atlasdb/cassandra/CassandraKeyValueServiceConfig.java
+++ b/atlasdb-cassandra/src/main/java/com/palantir/atlasdb/cassandra/CassandraKeyValueServiceConfig.java
@@ -329,11 +329,15 @@ default int socketQueryTimeoutMillis() {
      * When a Cassandra node is down or acting malignantly, it is plausible that we succeed in creating a socket but
      * subsequently do not read anything, and thus are at the mercy of the {@link #socketQueryTimeoutMillis()}. This is
      * particularly problematic on the creation of transaction managers and client pools in general: we can end up in
-     * a state where a query to a bad host blocks us for _at least_
+     * a state where a query to a bad host blocks us for up to 126 seconds with default settings (due to retrying three
+     * times on the first node).
+     *
+     * This initial timeout actually affects the creation of the clients themselves, and is only set for the call to
+     * login on Cassandra. When that is successful, the query timeout is set to the regular setting above.
      */
     @Value.Default
     default int initialSocketQueryTimeoutMillis() {
-        return 2 * 1000;
+        return 5 * 1000;
     }
 
     @Value.Default