I think the question is mainly around "how do we recognize the bookie" or "incarnations". And the purpose of a cookie is designed for addressing "incarnations".
I will try to cover following aspects, and will try to answer questions that Ivan and JV raised. - what is cookie? - how the behavior became bad? - how do we fix current bad behavior? - is the cookie enough? *What is Cookie?* Cookie is originally introduced in this commit - https://github.com/apache/bookkeeper/commit/c6cc7cca3a85603c8e935ba6d06fbf3d8d7a7eb5 . A cookie is a identifier of a bookie. A cookie is created on zookeeper when a brand new bookie joint the cluster, the cookie is representing the bookie instance during its lifecycle. The cookie is stored on all the disks for verification purpose. so if any of the disks misses the cookie (e.g. disks were reformat or wiped out, disks are not mounted correctly), a bookie will reject to start. *How the behavior became bad?* The original behavior worked as expected to use the cookie in zookeeper as the source of truth. See https://github.com/apache/bookkeeper/commit/c6cc7cca3a85603c8e935ba6d06fbf3d8d7a7eb5 The behavior was changed at https://github.com/apache/bookkeeper/commit/19b821c63b91293960041bca7b031614a109a7b8 when trying to support both ip and hostname . It used journal directory as the source-of-truth for verifying cookies. At the community meeting, I was saying a bookie should reject start when a cookie file is missing locally and that was my operational experience. It turns out twitter's branch didn't include the change at 19b821c63b91293960041bca7b031614a109a7b8, so it was still the original behavior at c6cc7cca3a85603c8e935ba6d06fbf3d8d7a7eb5 . *How do we fix current bad behavior?* We basically need to revert the current behaviour to the original designed behavior. The cookie in zookeeper should be the source-of-truth for validation. If the cookie works as expected (change the behavior to the original behavior), then it is the operational or lifecycle management issue I explained above. If a bookie failed with missing cookie, it should be: 1. taken out of the cluster 2. run re-replication (autorecovery or manual recovery) 3. ensure no ledgers using this bookie any more 4. reformat the bookie 5. add it back This can be automated by hooking into a scheduler (like k8s or mesos). But it requires some sort of lifecycle management in order to automate such operations. There is a BP-4: https://cwiki.apache.org/confluence/display/BOOKKEEPER/BP-4+-+BookKeeper+Lifecycle+Management proposed for this purpose. *Is the cookie enough?* Cookie (if we revert the current behavior to the original behavior), should be able to address most of the issues related to "incarnations". There are still some corner cases will violate correctness issues. They are related to "dangling writers" described in Ivan's first comment. How can a writer tell whether bookies changed or ledger changed when it gets network partitioned? 1) Bookie Changed. Bookie can be reformatted and re-added to the cluster. Ivan and JV already touch this on adding UUID. I think the UUID doesn't have to be part of ledger metadata. because auditor and replication worker would use the lifecycle management for managing the lifecycle of bookies. But the connection should have the UUID informations. Basically, any bookie client connects to a bookie, it needs to carry the namespace uuid and the bookie uuid to ensure bookie is connecting to a right bookie. This would prevent "dangling writers" connect to bookies that are reformatted and added back. 2) Ledger Changed. It is similar as what the case that Ivan' described. If a writer becomes 'network partitioned', and the ledger is deleted during this period, after the writer comes back, the writer can still successfully write entries to the bookies, because the ledgers are already deleted and all the fencing bits are gone. This violates the expectation of "fencing". but I am not sure we need to spend time on fixing this, because the ledger is already explicitly deleted by the application. so I think the behavior should be categorized as "undefined", just like "deleting a ledger when a writer is still writing entries" is a undefined behavior. To summarize my thought on this: 1. we need to revert the cookie behaviour to the original behavior. make sure the cookie works as expected. 2. introduce UUID or epoch in the cookie. client connection should carry namespace uuid and bookie uuid when establishing the connection. 3. work on BP-4 to have a complete lifecycle management to take bookie out and add bookie out. 1 is the immediate fix, so correct operations can still guarantee the correctness. - Sijie On Fri, Oct 6, 2017 at 9:35 AM, Venkateswara Rao Jujjuri <jujj...@gmail.com> wrote: > > However, imagine that the fenced message is only in the journal on b2, > > b2 crashes, something wipes the journal directory and then b2 comes > > back up. > > In this case what happened? > 1. We have WQ = 1 > 2. We had data loss (crash and comeup clean) > > But yeah, in addition to dataloss we have fencing violation too. > The problem is not just wiped journal dir, but how we recognize the bookie. > Bookie is just recognized by its ip address, not by its incarnation. > Bookie1 at T1 (b1t1) ; and same bookie1 at T2 after bookie format (b1t2) > should be two different bookies, isn;t it? > this is needed for the replication worker and the auditor too. > > Also, bookie needs to know if the writer/reader is intended to read from > b1t2 not from b1t1. > Looks like we have a hole here? Or I may not be fully understanding cookie > verification mechanism. > > Also as Ivan pointed out, we appear to think the lack of journal is > implicitly a new bookie, but overall cluster doesn't differentiate between > incarnations. > > Thanks, > JV > > > > > > On Fri, Oct 6, 2017 at 8:46 AM, Ivan Kelly <iv...@apache.org> wrote: > > > > The case you described here is "almost correct". But there is an key > > here: > > > B2 can't startup itself if journal disk is wiped out, because the > cookie > > is > > > missed. > > This is what I expected to see, but isn't the case. > > <snip> > > List<Cookie> journalCookies = Lists.newArrayList(); > > // try to read cookie from journal directory. > > for (File journalDirectory : journalDirectories) { > > try { > > Cookie journalCookie = > > Cookie.readFromDirectory(journalDirectory); > > journalCookies.add(journalCookie); > > if (journalCookie.isBookieHostCreatedFromIp()) { > > conf.setUseHostNameAsBookieID(false); > > } else { > > conf.setUseHostNameAsBookieID(true); > > } > > } catch (FileNotFoundException fnf) { > > newEnv = true; > > missedCookieDirs.add(journalDirectory); > > } > > } > > </snip> > > > > So if a journal is missing the cookie, newEnv is set to true. This > > disabled the later checks. > > > > > Hower it can still happen in a different case: bit flap. In your case, > if > > > fence bit in b2 is already persisted on disk, but it got corrupted. > Then > > it > > > will cause the issue you described. One problem is we don't have > checksum > > > on the index file header when it stores those fence bits. > > Yes, this is also an issue. > > > > -Ivan > > > > > > -- > Jvrao > --- > First they ignore you, then they laugh at you, then they fight you, then > you win. - Mahatma Gandhi >
