Process of vacuum and vacuum full

    For concurrency control of database system ,PostgreSQL use MVCC（ Multi version concurrency control ） To deal with . This mechanism has one drawback , As time goes on , Accumulated in data files dead tuples More and more . How to clean up these dead tuples, This is the time vacuum Handle .

    PostgreSQL Systematic vacuum It is a routine maintenance process , The system will also start when the service is started autovacuum The daemon maintains this , Of course, there are vacuum The command allows the user to manually execute vacuum operation . Except for cleaning up dead tuples,vacuum And frozen transactions ID The function of , Because now PostgreSQL The business of ID Only support 32 position （ Probably 40 A hundred million things ）, When a transaction ID After reaching the maximum value , It will re 0 Start , At this time, the previous affairs will become “ future ” The transaction becomes invisible .

    Here is the main introduction vacuum clear dead tuples The process of . Cleaning up dead tuples when , The system has two processing methods ：vacuum and vacuum full. The main differences are as follows ：

vacuum	vacuum full
Just put the... In the table dead tuples Make a deletion mark , There is no real physical deletion ;vacuum In the process , The data table can be accessed normally	Physically delete... In the table dead tuples, Free up space for the operating system ;vacuum full In the process , Table locked , Access not allowed

    Vacuum The process

    vacuum The implementation process of is mainly divided into the following three steps ：

    1、 eliminate dead tuples Point to the index tuples

    In the process ,vacuum Will scan the target table in sequence , And build a dead tuples Composed of list Linked list , The list The linked list will be stored in maintenance_work_mem In cache . then vacuum according to dead tuples list remove dead tuples Point to the index.

    2、 remove dead tuples, to update VM and FSM

    The removal here dead tuples Just mark the space as reusable , There is no real physical deletion . therefore vacuum After cleaning the table , The actual space of the table is not reduced .dead tuples After making the removal mark ,vacuum The remaining tuples are rearranged for defragmentation . then , Need to update the target table VM（ Visibility mapping file ） and FSM（ Free space mapping file ）.

    3、 Update statistics and related system tables

    Last , You need to update vacuum Statistics of the target table （ To adapt to the latest query optimization ） And related system tables .

    Vacuum Full The process

    Vacuum Full and Vacuum The big difference is this ,Vacuum Full It's physical deletion dead tuples, Free up space for the operating system , So in vacuum full after , The size of the table will be reduced to the actual space size . Its processing process and vacuum Very different , The processing steps are as follows ：

    1、vacuum full At the beginning of execution , The system will first create a AccessExclusiveLock, No more outside access is allowed （ Prepare for later copies ）, Then create a new table with the same table structure as the target table .

    2、 Scan target table , Put... In the watch live tuples Copy to new table .

    3、 Delete target table , On the new watch , Recreate index , to update VM,FSM And Statistics , Relevant system tables, etc .

    therefore ,vacuum full The essence of is to generate a new data file , Then put the original table live tuples Stored in the data file . contrast vacuum,vacuum full The disadvantage is that the table cannot be accessed during execution , Due to the need to import into the new table live tuples data , Its execution efficiency will also be very slow . The advantage is that after execution , Tablespaces only store live tuples, No redundancy dead tuples, It will improve the efficiency of executing queries .

    AutoVacuum Daemon

    because vacuum It is a routine maintenance of database system , So after the system starts , Will open a autovacuum Daemons deal specifically with this .autovacuum The cleaning process is generally used vacuum Avoid using vacuum full. The reason is that ,vacuum Although the minimum size of the table can be maintained , But if the table still has room to grow , It doesn't make much sense , Especially for frequently written tables . and vacuum It is also more efficient than vacuum full good .

    Above mentioned autovacuum The process is a launcher process , It doesn't do vacuum operation , It starts vacuum worker process , Then from vacuum worker The process proceeds accordingly vacuum and analyze operation .

    autovacuum The process will autovacuum_naptime Start the database within the time interval autovacuum worker process . Then each worker The process checks every table in the database , Execute when needed vacuum or analyze operation . If there is N A database , Then every autovacuum_naptime/N Seconds will start a new worker process , Only up to... Are allowed at one time autovacuum_max_workers individual worker Process is running , If there is more than autovacuum_max_workers Databases need to be processed , The next database will be in the first worker It will be handled immediately after completion .