diff --git a/src/bees-context.cc b/src/bees-context.cc
index a9722fe..153d7b5 100644
--- a/src/bees-context.cc
+++ b/src/bees-context.cc
@@ -40,30 +40,26 @@ BeesFdCache::BeesFdCache()
 	m_file_cache.max_size(BEES_FILE_FD_CACHE_SIZE);
 }
 
+void
+BeesFdCache::clear()
+{
+	BEESNOTE("Clearing root FD cache to enable subvol delete");
+	m_root_cache.clear();
+	BEESCOUNT(root_clear);
+	BEESNOTE("Clearing open FD cache to enable file delete");
+	m_file_cache.clear();
+	BEESCOUNT(open_clear);
+}
+
 Fd
 BeesFdCache::open_root(shared_ptr<BeesContext> ctx, uint64_t root)
 {
-	// Don't hold root FDs open too long.
-	// The open FDs prevent snapshots from being deleted.
-	// cleaner_kthread just keeps skipping over the open dir and all its children.
-	if (m_root_cache_timer.age() > BEES_COMMIT_INTERVAL) {
-		BEESLOGINFO("Clearing root FD cache to enable subvol delete");
-		m_root_cache.clear();
-		m_root_cache_timer.reset();
-		BEESCOUNT(root_clear);
-	}
 	return m_root_cache(ctx, root);
 }
 
 Fd
 BeesFdCache::open_root_ino(shared_ptr<BeesContext> ctx, uint64_t root, uint64_t ino)
 {
-	if (m_file_cache_timer.age() > BEES_COMMIT_INTERVAL) {
-		BEESLOGINFO("Clearing open FD cache to enable file delete");
-		m_file_cache.clear();
-		m_file_cache_timer.reset();
-		BEESCOUNT(open_clear);
-	}
 	return m_file_cache(ctx, root, ino);
 }
 
diff --git a/src/bees-roots.cc b/src/bees-roots.cc
index 863378d..0b4c7d7 100644
--- a/src/bees-roots.cc
+++ b/src/bees-roots.cc
@@ -335,11 +335,22 @@ BeesRoots::crawl_thread()
 
 	// Monitor transid_max and wake up roots when it changes
 	BEESNOTE("tracking transids");
+	auto last_count = m_transid_re.count();
 	while (true) {
 		// Make sure we have a full complement of crawlers
 		// Calls transid_max() which updates m_transid_re
 		insert_new_crawl();
 
+		// Don't hold root FDs open too long.
+		// The open FDs prevent snapshots from being deleted.
+		// cleaner_kthread just keeps skipping over the open dir and all its children.
+		// Even open files are a problem if they're big enough.
+		auto new_count = m_transid_re.count();
+		if (new_count != last_count) {
+			m_ctx->fd_cache()->clear();
+		}
+		last_count = new_count;
+
 		BEESNOTE("waiting for next transid " << m_transid_re);
 		// We don't use wait_for here because somebody needs to
 		// be updating m_transid_re from time to time.
diff --git a/src/bees.h b/src/bees.h
index 38d51c4..1b416e8 100644
--- a/src/bees.h
+++ b/src/bees.h
@@ -59,11 +59,8 @@ const off_t BLOCK_SIZE_HASHTAB_EXTENT = 16 * 1024 * 1024;
 // Bytes per second we want to flush (8GB every two hours)
 const double BEES_FLUSH_RATE = 8.0 * 1024 * 1024 * 1024 / 7200.0;
 
-// How long we should wait for new btrfs transactions
-const double BEES_COMMIT_INTERVAL = 900;
-
-// Interval between writing non-hash-table things to disk, and starting new subvol crawlers
-const int BEES_WRITEBACK_INTERVAL = BEES_COMMIT_INTERVAL;
+// Interval between writing crawl state to disk
+const int BEES_WRITEBACK_INTERVAL = 900;
 
 // Statistics reports while scanning
 const int BEES_STATS_INTERVAL = 3600;
@@ -92,7 +89,7 @@ const double BEES_TOO_LONG = 5.0;
 // Avoid any extent where LOGICAL_INO takes this long
 // const double BEES_TOXIC_DURATION = 9.9;
 // EXPERIMENT:  Kernel v4.14+ may let us ignore toxicity
-const double BEES_TOXIC_DURATION = BEES_COMMIT_INTERVAL;
+const double BEES_TOXIC_DURATION = 99.9;
 
 // How long between hash table histograms
 const double BEES_HASH_TABLE_ANALYZE_INTERVAL = BEES_STATS_INTERVAL;
@@ -657,6 +654,7 @@ public:
 	Fd open_root(shared_ptr<BeesContext> ctx, uint64_t root);
 	Fd open_root_ino(shared_ptr<BeesContext> ctx, uint64_t root, uint64_t ino);
 	void insert_root_ino(shared_ptr<BeesContext> ctx, Fd fd);
+	void clear();
 };
 
 struct BeesResolveAddrResult {