I've implemented PeersContainer class with all operations that I need. I need something like multimap in concurrent environment, moreover I need remove entities automatically on timeout. So, I decide to use Cache<Long, BlockingDeque<Peer>> combined with Striped<ReadWriteLock>. Also I use two caches which can transfer entities between each other.

Question: Is all those operations thread-safe in my context?

If it isn't thread safe, please, provide an example which demonstrate scenario where I'll get inconsistent data.

public class PeersContainer {

private final static int MAX_CACHE_SIZE = 100;
private final static int STRIPES_AMOUNT = 10;
private final static int PEER_ACCESS_TIMEOUT_MIN = 30;

private final static PeersContainer INSTANCE;

static {
INSTANCE = new PeersContainer();
}

public static PeersContainer getInstance() {
return INSTANCE;
}

private final Cache<Long, BlockingDeque<Peer>> peers = CacheBuilder.newBuilder()
.maximumSize(MAX_CACHE_SIZE)
.expireAfterAccess(PEER_ACCESS_TIMEOUT_MIN, TimeUnit.MINUTES)
.removalListener(new RemovalListener<Long, BlockingDeque<Peer>>() {
public void onRemoval(RemovalNotification<Long, BlockingDeque<Peer>> removal) {
if (removal.getCause() == RemovalCause.EXPIRED) {
//send logOut response to all use'r {@code Peer} objects.
for (Peer peer : removal.getValue()) {
//peer.sendLogOutResponse();
}
}
}
})
.build();

private final Cache<Long, UserAuthorities> authToRestore = CacheBuilder.newBuilder()
.maximumSize(MAX_CACHE_SIZE)
.expireAfterWrite(PEER_ACCESS_TIMEOUT_MIN, TimeUnit.MINUTES)
.build();

public BlockingDeque<Peer> getPeers(long sessionId) {
peers.cleanUp();
lock.lock();
try {
return peers.asMap().get(sessionId);
} finally {
lock.unlock();
}
}

long key = peer.getSessionId();
Lock lock = stripes.get(key).writeLock();
lock.lock();
try {
BlockingDeque<Peer> userPeers = peers.getIfPresent(key);
if (userPeers == null) {
peers.put(key, userPeers);
}
UserAuthorities authorities = restoreSession(key);
if (authorities != null) {
peer.setAuthorities(authorities);
}
return userPeers.offer(peer);
} finally {
lock.unlock();
}
}

public void removePeer(Peer peer) {
long sessionId = peer.getSessionId();
Lock lock = stripes.get(sessionId).writeLock();
peers.cleanUp();
lock.lock();
try {
UserAuthorities authorities = null;
if (peers.getIfPresent(sessionId) != null) {
authorities = peers.getIfPresent(sessionId).getFirst().getAuthorities();
authToRestore.put(sessionId, authorities);
peers.getIfPresent(sessionId).remove(peer);
}
} finally {
lock.unlock();
}
}

public void removePeers(long sessionId) {
Lock lock = stripes.get(sessionId).writeLock();
lock.lock();
try {
peers.invalidate(sessionId);
authToRestore.invalidate(sessionId);
} finally {
lock.unlock();
}
}

private UserAuthorities restoreSession(long sessionId) {
peers.cleanUp();
authToRestore.cleanUp();
lock.lock();
try {
BlockingDeque<Peer> activePeers = peers.getIfPresent(sessionId);
return (activePeers != null && !activePeers.isEmpty()) ? activePeers.getFirst().getAuthorities() : authToRestore.getIfPresent(sessionId);
} finally {
lock.unlock();
}
}
}

-

From a concurrency stand-point, I don't see any issue with the code: reading is guarded by the readLock, all the updates are guarded by the writeLock, and you're using thread-safe BlockingDeques for the values.

However, you're returning the actual BlockingDeque<Peer> from the Cache in getPeers(), so you're not mandating going through your PeersContainer to manipulate the values. You depend on how the other objects collaborate.

# 2. Other considerations

From a more general stand-point, I'd make a few changes:

## 1. Consistent API usage

public BlockingDeque<Peer> getPeers(long sessionId) {
peers.cleanUp();
lock.lock();
try {
return peers.asMap().get(sessionId);
} finally {
lock.unlock();
}
}


Why use asMap()? I'd stay with :

return peers.getIfPresent(sessionId);


## 2. Intermixing blocks

Especially with locks, I like them close to where I use them. So instead of

Lock lock = stripes.get(sessionId).readLock();
peers.cleanUp();
lock.lock();


I'd write

peers.cleanUp();
lock.lock();


## 3. Multiple look-ups

Even if Cache is fast, no need to abuse it :). In removePeer():

UserAuthorities authorities = null;
if (peers.getIfPresent(sessionId) != null) {
authorities = peers.getIfPresent(sessionId).getFirst().getAuthorities();
authToRestore.put(sessionId, authorities);
peers.getIfPresent(sessionId).remove(peer);
}


can be replaced by (also fixing the unconditional getFirst() call):

BlockingDeque<Peer> userPeers = peers.getIfPresent(sessionId);
if (userPeers != null && !userPeers.isEmpty()) {
UserAuthorities authorities = userPeers.getFirst().getAuthorities();
authToRestore.put(sessionId, authorities);
userPeers.remove(peer);
}


## 4. Excessive cleanUp

I understand you don't want to wait for Guava to decide when the removals should actually happen, but forcing it on every access might be overkill (and create more contention), and you still depend on a call to PeersContainer to trigger the clean-up.

You could use some form of scheduler to call a global PeersContainer.cleanUp() method every minute, instead.

## Unencapsulated modifications

By returning the actual BlockingDeque<Peer> from the Cache in getPeers(), you leave the possibility to modify the shared collection outside of the PeersContainer context, which may or not have an impact depending on the "friendliness" of the code using that data. If it's under your control and you never modify (even by accident) the collection, then it should be fine.

Otherwise, there are obviously 2 types of modifications:

• removing a Peer: OK, it's not there anymore, but to the PeersContainer it's not that big a deal, it can also happen if the entry expires in the cache
• adding a Peer: it won't have its authorities set, which means it might erase a UserAuthorities saved in the authToRestore cache by the removal of a previous Peer for the same session. Is it critical? I don't know, it's your application.

## Cache expiration

From the CacheBuilder javadoc:

If expireAfterWrite or expireAfterAccess is requested entries may be evicted on each cache modification, on occasional cache accesses, or on calls to Cache.cleanUp(). Expired entries may be counted by Cache.size(), but will never be visible to read or write operations.

Even if your cache doesn't have enough write activity, with enough read activity the expired entries will eventually be removed. But even if they're not removed yet, the Cache will never return entries once they're expired (it wouldn't be doing a very good job otherwise).

The only reason why you may need to call cleanUp() in your case is because of the nature of your RemovalListener: you want to send some information back to the peer, so you want to do it in a timely manner after the entry has expired. It's up to you to define what "timely" means depending on your requirements, but in my opinion, it could be something between 15 seconds and 5 minutes.

Anyway, the way it's implemented now, if you have a single, inactive peer (so no calls at all to the methods of PeersContainer), it'll never receive a logout, whereas with a scheduled task it would happen no matter what. And on the other hand, if you have enough activity in the application, the clean-up will happen naturally with the writes or numerous reads, and an extra, scheduled cleanUp call won't cost much, whereas calling cleanUp() on each and every call will cause contention because the clean-up needs to sequentially lock all the segments in the Cache to clean them one by one.

Another round.

You don't really need a ReadWriteLock:

• the readLock in getPeers() is useless since the Cache itself is thread-safe
• the one in restoreSession() is also useless because it's already guarded by the writeLock in addPeer(), its caller.

So you could simplify a bit by using a simple Lock. The only parts you really need to guard by a lock are the ones where you modify the 2 Cachess.

-
thanks for broad response. 1) Outside the PeerContainer I should only iterate the collection, that's why a implemented get method. Of course I can return a copy of BlocckingDeque, but I can't imagine scenario when it's harmful. Can you help? 4) As I understand Cache proceeds cleanups only on write operations. Then to force cache to return actual data on read operations I should cleanup explicitly. Am I wrong? – Volodymyr Bakhmatiuk Mar 3 '14 at 15:20
Answer updated. – Frank Pavageau Mar 3 '14 at 17:12
I've gotten a new insight after reading your other SO questions and sleeping on it: answer updated again. – Frank Pavageau Mar 4 '14 at 8:55
thanks I'll do that. But I still can't grasp how the second cache will be helpfull for me/ – Volodymyr Bakhmatiuk Mar 4 '14 at 9:23
By "the 2 caches", I meant peers and authToRestore, not the extra cache in stackoverflow.com/a/22155141/1350869. – Frank Pavageau Mar 4 '14 at 9:32