Postgres lock behaves inconsistently over different goroutines

Question

I've encountered some inconsistent behaviour with pg_locks on go routines and PostgreSQL 9.5.

When I create additional goroutines and call SELECT pg_try_advisory_lock(1); the results are inconsistent.

I can try to get the lock at some moment, fail, try again and manage to get it, without anyone explicitly releasing the lock in the meanwhile.

I've created the a small program to reproduce the issue.

Program flow

1. Create 10 goroutines. Each one of them tries to get the same lock on initialization.
2. Every second, each instance would try to get the lock again, if it didn't get it already.
3. Every second I probe all instances and count how many already got the lock.

Expected behaviour:

Only 1 goroutine would have the lock at any given moment.

Actual results:

The number of goroutines that manage to aquire the lock increases over time.

---

package main

var dbMap *gorp.DbMap // init code omitted for brevity

func main() {
    setup(10)
    fmt.Println("after initialization,", countOwners(), "instances of lockOwners have the lock!")

    for {
        if _, err := dbMap.Exec("SELECT pg_sleep(1)"); err != nil {
            panic(err)
        }

        fmt.Println(countOwners(), "instances of lockOwners have the lock!")
    }
}

func countOwners() int {
    possessLock := 0
    for _, lo := range los {
        if lo.hasLock {
            possessLock++
        }
    }
    return possessLock
}

var los []*lockOwner

func setup(instanceCount int) {
    var wg sync.WaitGroup
    for i := 0; i < instanceCount; i++ {
        wg.Add(1)
        newInstance := lockOwner{}
        los = append(los, &newInstance)
        go newInstance.begin(time.Second, &wg, i+1)
    }
    wg.Wait()
}

type lockOwner struct {
    id      int
    ticker  *time.Ticker
    hasLock bool
}

func (lo *lockOwner) begin(interval time.Duration, wg *sync.WaitGroup, id int) {
    lo.ticker = time.NewTicker(interval)
    lo.id = id
    go func() {
        lo.tryToGetLock()
        wg.Done()
        for range lo.ticker.C {
            lo.tryToGetLock()
        }
    }()
}

func (lo *lockOwner) tryToGetLock() {

    if lo.hasLock {
        return
    }

    locked, err := dbMap.SelectStr("SELECT pg_try_advisory_lock(4);")
    if err != nil {
        panic(err)
    }

    if locked == "true" {
        fmt.Println(lo.id, "Did get lock!")
        lo.hasLock = true
    }
}

The output of this program varies, but usually something along the lines:

1 Did get lock!
after initialization, 1 instances of lockOwners have the lock!
1 instances of lockOwners have the lock!
2 Did get lock!
2 instances of lockOwners have the lock!
2 instances of lockOwners have the lock!
7 Did get lock!
3 instances of lockOwners have the lock!
3 instances of lockOwners have the lock!
6 Did get lock!
4 instances of lockOwners have the lock!

---

My question:

1. What should I expect to be protected when using pg_locks this way?
2. What is the reason some goroutine fails to acquire the lock?

3. What is the reason the same goroutine succeeds doing so on the next attempt?

   • Could it be that the thread is the resource being locked and every time a goroutine triggers it does so from a different thread? That would explain the inconsistent behaviour.

Answer 1

After some months of experience with PostgreSQL via gorp, I think I understand this behaviour:

• gorp maintains a connection pool.
• Whenever we create a transaction, one of these connections will be picked in random(?).
• dbMap.SomeCommand() creates a transaction, executes some command and commits the transaction.
• pg_try_advisory_lock works on the session level.
• A session is synonymous with a TCP connection, so is by no means stable or permanent.
• A connection from the pool keeps using the same session when possible, but will reset it when needed.

Whenever we execute dbMap.SelectStr("SELECT pg_try_advisory_lock(4);"), a connection is selected from the pool. Then a transaction is created, it acquires the lock on the session level and then it's committed.

When another go-routine attempts to do the same, there's a chance that it will use the same session, probably depending on the connection that is taken from the pool. Since the locking is done on the session level - the new transaction is able to acquire the lock again.