diff --git a/lib/mastodon/sidekiq_middleware.rb b/lib/mastodon/sidekiq_middleware.rb
index 3a747afb63..ac63d4d38c 100644
--- a/lib/mastodon/sidekiq_middleware.rb
+++ b/lib/mastodon/sidekiq_middleware.rb
@@ -8,6 +8,7 @@ class Mastodon::SidekiqMiddleware
   rescue Mastodon::HostValidationError
     # Do not retry
   rescue => e
+    clean_up_elasticsearch_connections!
     limit_backtrace_and_raise(e)
   ensure
     clean_up_sockets!
@@ -25,6 +26,32 @@ class Mastodon::SidekiqMiddleware
     clean_up_statsd_socket!
   end
 
+  # This is a hack to immediately free up unused Elasticsearch connections.
+  #
+  # Indeed, Chewy creates one `Elasticsearch::Client` instance per thread,
+  # and each such client manages its long-lasting connection to
+  # Elasticsearch.
+  #
+  # As far as I know, neither `chewy`,  `elasticsearch-transport` or even
+  # `faraday` provide a reliable way to immediately close a connection, and
+  # rely on the underlying object to be garbage-collected instead.
+  #
+  # Furthermore, `sidekiq` creates a new thread each time a job throws an
+  # exception, meaning that each failure will create a new connection, and
+  # the old one will only be closed on full garbage collection.
+  def clean_up_elasticsearch_connections!
+    return unless Chewy.enabled? && Chewy.current[:chewy_client].present?
+
+    Chewy.client.transport.connections.each do |connection|
+      # NOTE: This bit of code is tailored for the HTTPClient Faraday adapter
+      connection.connection.app.instance_variable_get(:@client)&.reset_all
+    end
+
+    Chewy.current.delete(:chewy_client)
+  rescue
+    nil
+  end
+
   def clean_up_redis_socket!
     RedisConfiguration.pool.checkin if Thread.current[:redis]
     Thread.current[:redis] = nil