elementsd shutdown seems unnecessarily slow for long-running nodes

まず、ブロックインデックスを手動でアンロードし、そのエントリを解放するシャットダウン経路を特定し、次にそれをプロセス終了時の動作と比較します。可能であれば、長時間稼働した node の後にシャットダウンを再現し、複数のブロックインデックスを作成および破棄するテストケースを調査します。テストで必要とされる場合のクリーンな削除を維持しながら、不要なシャットダウン処理を避けられれば完了です。

I'm curious if other people are finding that elementsd shutdown takes forever when the node has been running for a very long time, or if this is somehow unique to my machine. It's possible this issue is somehow specific to mac or non-linux platforms; it's also possible it's been made worse by my memory-reducing fix.

What appears to be happening: during shutdown we manually unload the block index, and deallocate the entries, one at a time. This seems basically pointless -- all we're doing is manually freeing memory that will be freed anyway when the process exits. This does not seem to be slow if we've just started, but takes forever when run on a process that's been around for a long time. What I suspect is happening is that the block index has been paged out to disk, and in clearing it we're paging it in, a bit at a time, as we access the entries in random order. Even at SSD speeds, paging in multiple gigabytes using random reads is going to take ages.

(The more problematic possibility is some kind of leak, but I don't see the process having particularly excessive memory usage -- any more than usual -- so I don't think this is it.)

If possible, it would be nice to stop wasting all this effort on shutdown. Assuming it's in fact wasted, one possible catch is the use of the same codepaths in test cases -- we may still need to cleanly delete objects there, if we have to create and destroy multiple block indices before ending the process.