Proposal For Read-After-Write Performance Improvement

Introduction

Hi, I find this project very cool and want to be a contributor, I write this proposal based on the ReadAfterWrite Consistency Document, and the changes I make are marked in Bold or Delete Line.

This is only a preliminary version, and I hope that I can fully discuss the optimization logic with community members before designing the code implementation, looking forward to your reply : )

Goals

• Session Level ReadAfterWrite: Ensure read requests get latest write in the same client connection.
• Instance Level ReadAfterWrite: Ensure read requests get latest write in the WeSQL WeScale Instance.
• Implement Table-Level Read-After-Write Support.

Design Details

Step 1: Get GTID after write operation without extra network round

Starting from MySQL 5.7, the MySQL protocol implements a mechanism to collect the GTIDs to be sent over the wire in the response packet. This feature assists us in acquiring GTIDs without introducing further network rounds.

To enable the feature:

• Client needs to set Capability flag CLIENT_SESSION_TRACK when connecting to MySQL via mysql protocol. This will enable MySQL to send the tracking information back to the client.
• Client also needs to issue SET @@SESSION_TRACK_GTIDS = 'OWN_GTID' to indicate MySQL to return GTID in the OK packet. This system variable tracks the last DML and DDL commit GTID.

Step 2: Manage the latest GTID and update time for each table in the last t senonds

We can use struct LatestGTIDManager to manage the latest GTID and update time for each table.

The code below is just used to illustrate the method:

// LatestGTIDEntry represents an entry in the LatestGTIDManager with the table name, GTID, and the time it was updated.  
type LatestGTIDEntry struct {  
	GTID     string  
	UpdateTime time.Time  
}  
  
// LatestGTIDManager manages the latest GTID and update time for each table.  
type LatestGTIDManager struct {  
	latestGTIDs map[string]LatestGTIDEntry // Key is the table name, value is the LatestGTIDEntry struct.  
	expireTime  time.Duration               // The expiration time for GTID entries.  
	mu          sync.RWMutex               // Mutex for read-write synchronization.  
	wg          sync.WaitGroup             // WaitGroup to wait for the cleanup goroutine to finish.  
}  
  
// NewLatestGTIDManager creates a new instance of LatestGTIDManager.  
func NewLatestGTIDManager(expireTime time.Duration) *LatestGTIDManager {  
	return &LatestGTIDManager{  
		latestGTIDs: make(map[string]LatestGTIDEntry),  
		expireTime:  expireTime,  
	}  
}  
  
// UpdateGTID updates the latest GTID and update time for a given table.  
func (m *LatestGTIDManager) UpdateGTID(tableName, gtid string) {  
	m.mu.Lock()  
	defer m.mu.Unlock()  
	m.latestGTIDs[tableName] = LatestGTIDEntry{  
		GTID:     gtid,  
		UpdateTime: time.Now(),  
	}  
}  
  
// GetLatestGTID retrieves the latest GTID for a given table.  
// If the table is not found or the GTID has expired, it returns an empty string and false.  
func (m *LatestGTIDManager) GetLatestGTID(tableName string) (string, bool) {  
	m.mu.RLock()  
	defer m.mu.RUnlock()  
	entry, ok := m.latestGTIDs[tableName]  
	if !ok || time.Now().Sub(entry.UpdateTime) > m.expireTime {  
		return "", false  
	}  
	return entry.GTID, true  
}  
  
// startCleaner starts a goroutine to periodically clean up expired GTID entries.  
func (m *LatestGTIDManager) startCleaner() {  
	m.wg.Add(1)  
	go func() {  
		defer m.wg.Done()  
		ticker := time.NewTicker(m.expireTime)  
		defer ticker.Stop()  
		for {  
			select {  
			case <-ticker.C:  
				m.mu.Lock()  
				now := time.Now()  
				for tableName, entry := range m.latestGTIDs {  
					if now.Sub(entry.UpdateTime) > m.expireTime {  
						delete(m.latestGTIDs, tableName)  
					}  
				}  
				m.mu.Unlock()  
			}  
		}  
	}()  
}  
  
// Stop waits for the cleanup goroutine to finish.  
func (m *LatestGTIDManager) Stop() {  
	m.wg.Wait()  
}  

Depends on the consistency level, the LatestGTIDManager may be initialized in the client’s session or a global memory data structure.

// Initialize LatestGTIDManager with an expiration time of 10 minutes.  
gm := NewLatestGTIDManager(10 * time.Second)  
gm.startCleaner()

Step 3: Store the GTID in WeSQL WeScale sessions

After parsing the response packet and get the GTIDs, WeSQL WeScale will store them in the memory.

If the operation is a write operation, the LatestGTIDManager will update the latest GTID and write time for the table has be written.

gm.UpdateGTID("my_table", "abcdefg-1234567-890") 

Depends on the consistency level, the GTIDs may be stored in the client’s Session or a global memory data structure.

When a read operation happens, we will utilize the LatestGTIDManager to get the Latest_GTID_for_Table_to_be_Read.

Latest_GTID_for_Table_to_be_Read, ok := gm.GetLatestGTID("my_table") 

Two situations will occur at this time:

• The table has been updated in the last t seconds, we get its Latest_GTID_for_Table_to_be_Read, and enter the Step 4.
• The table has NOT been updated in the last t seconds, the information for this table has been cleaned up by LatestGTIDManager, at this point, in a radical way, since the last write to the table has been at least t seconds, it can be considered that the last write to the table has been completed from every follower, we can just pick a follower and read it.

Later read operations will utilize GTIDs stored in WeSQL WeScale’s memory, to ensure retrieval of data that was previously written. See belowing steps for more details.

Step 4: Select a MySQL follower for reading

A CLUSTER_GTID_EXEUTED memory data structure is matained in WeSQL WeScale’s memory, it contains’s all the @@global.gtid_executed values from the cluster. The CLUSTER_GTID_EXEUTED is updated by the health-check module periodically, and obviously it will be lagging.

Therefore, GTIDs from step1 will update CLUSTER_GTID_EXEUTED constantly.

During routing phase of a read operation, it will use the GTID (from session or global memory data structure) to pick a MySQL instance based on CLUSTER_GTID_EXEUTED.

During routing phase of a read operation, it will use the Latest_GTID_for_Table_to_be_Read (from LatestGTIDManager stored in session or global memory) to pick a MySQL instance based on CLUSTER_GTID_EXEUTED.

As long as the picked MySQL instance containes the Latest_GTID_for_Table_to_be_Read, the read operation can be directly forwarded to the MySQL instance.

Step 5: Ensure write requests have been propagated to the follower MySQL

All the follower MySQL instances may be lagging, or the CLUSTER_GTID_EXEUTED may be out-of-date for whatever reasons. It is possible that no follower (expect leader, it always holds all the data) is available for a read operation in Step 4.

We can either send the read operation to the leader, or send the read operation to the follower with a WAIT_FOR_EXECUTED_GTID_SET prefix. WAIT_FOR_EXECUTED_GTID_SET function will keep waiting until a GTID is executed on the follower or until times out.

We can use multi-statement to save one network round:

-- for example, if user's SQL is: select * from t1;

-- the actual SQL sent to follower may be a multi-statement like this:
select WAIT_FOR_EXECUTED_GTID_SET('ab73d556-cd43-11ed-9608-6967c6ac0b32:7', 3);select * from t1;

We need to handle the mysql protocol carefully to use the multi-statement, otherwise the mysql connection may be broken.

Thank you for your interest in this topic. If you would like to proceed, please feel free to send an email to 399geray@gmail.com.
Your understanding is correct, and we can discuss the implementation details further. We should consider the scalability of the implementation because this feature fundamentally analyzes the dependency between two SQL. The basic approach is at the table level, and we will implement a more fine-grained dependency detection.

Cool! I have sent you an email about the idea of dependency detection , I'll take some more time to read the source code carefully and look forward to discussing it with you further!

Hi, I am very interested in this issue. Yesterday, I sent an email outlining some of my thoughts and ideas. I look forward to the opportunity to discuss them with you further.Thank you for your time and consideration!