apple/foundationdb

DD: Build teams in batches to avoid slow task from DD

Open

#2,102 opened on 2019年9月12日

GitHub で見る
 (3 comments) (0 reactions) (0 assignees)C++ (13,912 stars) (1,266 forks)batch import
data distributiongood first issue

説明

The current DD build all required teams in one big atomic block.

When there are a lot of teams to build, say over 6000 teams, DD will take tens of seconds (say 20 seconds) to build all teams without yield. This leads to slow task which makes DD non-responsive to other requests, such as heart beat. Slow tasks can lead to problems: such as cluster controller can mark a live DD as dead, and data balance can be blocked.

The solution to this issue is to break the non-stopping block of buildTeams() into smaller asynchronous blocks. When a lot of teams are to build, buildTeams() will build some of them (say 100 teams) and yield. Later, buildTeams() build another batch of teams.

This issue is related to Issue #2096.

コントリビューターガイド

技術スタック
cpp
領域
databaseperformance
Issue 種別
performance
難度新しい貢献者にとっての実装難度です。1 は非常に小さな変更、5 は専門的な作業です。
3
推定時間経験ある貢献者が調査、実装、テスト、pull request 準備にかけるおおよその時間範囲です。
half day
活動状況Issue が今どれだけ取り組みやすいかを示します。新しい、活発、古い、ブロック中、またはメンテナー入力待ちなどです。
stale
明確さ期待される変更、受け入れ条件、次の手順がどれだけ明確かを示します。
clear
前提条件
C++ programmingFoundationDB architectureasynchronous programming concepts
初心者向け度初回貢献者にどれだけ取り組みやすいかを 1-100 で推定したスコアです。
35
調査方針
Examine the buildTeams() function in the Data Distributor (DD) component. Look at how teams are currently built atomically. Consider splitting the work into batches of 100 teams using asynchronous operations. Check issue #2096 for related context. Ensure that the batching does not introduce data races or deadlocks. Test with large numbers of teams to verify performance improvement.
DD: Build teams in batches to avoid slow task from DD · apple/foundationdb#2102 | Good First Issue