Mysql zero downtime deployment

Why we cannot perform an alter table directly to update schema?

How mysql performs alter table:

  1.  Lock the table

  2. Make a copy of the table 

  3. Modify the copy (the "new table")

  4. Copy all the rows into the new table

  5. Swap the old and new table

  6. Unlock the table

The key issue here is that lock time will be quite long for large table, whick blocks online transaction.


Percona's solution:

  • pt-online-schema-change tool:

    How pt-online-schema-change performs alter table?

  1. Make a copy of the table

  2. Modify the copy (the "new table")

  3. Copy all the rows into the new table (do in small chunks,  insert .. select)

  4. Add triggers to keep track of changes

  5. Swap the old and new table

Pros:
     Remove the Lock/unlock steps, no longer blocking
     Replication-Awareness
     Load awareness -> chunk size is auto adjusted

Cons:
     About 4th slower than alter table directly

Limitation:
     Cannot handle foreign key perfectly

  • XTraBackup:

  1. Detect and write redo log changes to xtrabackup_log

  2. Copy .ibd; idbdata1 to backup folder with timestamp

  3. Flush table with read lock

  4. Copy .frm  .myd  .myi files

  5. Get binary log position

  6. Unlock tables

  7. Stop and copy xtrabackup_log


mysqldump --single-transaction  --master-data


Reference:

https://www.percona.com/files/presentations/WEBINAR-zero-downtime-schema-changes.pdf 

http://www.cnblogs.com/olinux/p/5207887.html

http://www.cnblogs.com/cchust/p/4603599.html


 



 

發表評論
所有評論
還沒有人評論,想成為第一個評論的人麼? 請在上方評論欄輸入並且點擊發布.
相關文章