com.sleepycat.je.rep.impl.node.cbvlsn

Class GlobalCBVLSN

java.lang.Object

com.sleepycat.je.rep.impl.node.cbvlsn.GlobalCBVLSN

public class GlobalCBVLSN
extends Object

The GlobalCBVLSN was used prior to JE 7.5 for ensuring that nodes do not delete files needed for feeding other nodes. The methods in this class are disabled if the GlobalCBVLSN is defunct -- see below.

In version 7.5 and greater we instead retain all files up to the disk limits just in case they are needed for replication, and the GlobalCBVLSN is no longer needed. The FileProtector is used for preventing file deletion on a local node. See [#25220] in the JE 7.5 change log for a list of related external changes.

The global and local CBVLSNs are completely defunct (no longer maintained) in a rep group where all nodes are running JE 7.5 or greater. In a mixed version group of new (7.5 or later) and old (pre-7.5) nodes, we must continue to maintain the GlobalCBVLSN for the sake of the old nodes. If we did not do this, for example, the following scenario could occur:

  - RF-3, N1 is new, N2 and N3 are old.
  - N1 is the master, N2 is an up-to-date replica, N3 is a lagging replica.
  - N2 deletes files that would be needed to feed N3, because the GlobalVLSN
    is not updated and therefore N2 doesn't know that N3 is lagging.
  - N1 goes down, N2 is elected master
  - N3 requires a network restore because N2 doesn't have the files that N3
    needs for syncup.

The c.s.je.rep.impl.node.cbvlsn package contains most of the code for maintaining the CBVLSN. This code is isolated here for two reasons:

• To make it clear that it is not in play for groups of all new nodes. We don't have to think about the CBVLSN at all in this case.
• To make it easier to eventually remove this code entirely, when backward compatibility with JE 7.4 and earlier is dropped.

When it is not yet defunct, the GlobalCBVLSN class represents this node's view of the global CBVLSN. Each node has its own view of the global CBVLSN, based upon its local replicated copy of the rep group db. There is a single instance of the GlobalCBVLSN and it exists for the lifetime of the RepNode.

A global CBVLSN is a per-environment value, and is safeguarded from decreasing during the lifetime of a single RepImpl. Because nodes crash and re-sync, and new nodes join, it's possible that the persisted local cbvlsns can decrease, and drag the global cbvlsn value down, but those decreases are ignored during the lifetime of the global CBVLSN instance.

On pre-7.5 nodes, the global CBVLSN is used by:

• As a barrier to file deletion by the Cleaner.
• The Feeder which only serves log records in the interval: [GlobalCBVLSN .. VLSNRange.last]
• The Replica which uses the interval [GlobalCBVLSN .. VLSNRange.last] at syncup time.

Constructor Summary

Constructors

Constructor and Description
GlobalCBVLSN(RepNode repNode)

Method Summary

Modifier and Type	Method and Description
VLSN	getCBVLSN() Returns the global CBVLSN if it is not defunct.
VLSN	getRestoreResponseVLSN(VLSNRange range) Returns a VLSN appropriate for the RestoreResponse.cbvlsn field when the GlobalCBVLSN is not defunct.
void	init(RepNode repNode, JEVersion minJEVersion) Uses minJEVersion to determine whether the global CBVLSN is defunct.
boolean	isDefunct() Returns true if all nodes in the group are running JE 7.5 or later, and therefore the GlobalCBVLSN is not maintained.
void	recalculate(RepGroupImpl groupInfo) Recalculates the GlobalVLSN when it is not defunct.
void	setDefunctJEVersion(RepNode repNode) Calls RepNode.setMinJEVersion to upgrade the min version to DEFUNCT_JE_VERSION.
void	updateGroupInfo(NameIdPair updateNameIdPair, CleanerBarrierState barrierState) Updates the cached group info for the node, avoiding a database read, if the global CBVLSN is not defunct.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

GlobalCBVLSN

public GlobalCBVLSN(RepNode repNode)

Method Detail

init

public void init(RepNode repNode,
                 JEVersion minJEVersion)

Uses minJEVersion to determine whether the global CBVLSN is defunct. Try to avoid any use of the GlobalCBVLSN when all nodes are running new software: - For the master, this method should be called using the minJEVersion in the rep group DB. - For a replica, it should be called using the minJEVersion of the feeder, from the feeder handshake. See setDefunctJEVersion(RepNode) for how the minJEVersion is set for a new group (with new software) and for an upgraded group.

Parameters:

minJEVersion - the min JE version for the group, or null if unknown.

setDefunctJEVersion

public void setDefunctJEVersion(RepNode repNode)

Calls RepNode.setMinJEVersion to upgrade the min version to DEFUNCT_JE_VERSION. This allows a group to treat the global CBLVSN as defunct when all the electable nodes in the group have been upgraded. This method should be called when the first node in a a new group is created to avoid all use of the GlobalVLSN. It should also be called periodically on the master, as feeders become active, to update the min JE version after all nodes in a group are upgraded. If the version update fails because not all nodes are upgraded, or there is no quorum for the group DB update, or there is a lock conflict, this method fails silently and should be called again later.

isDefunct

public boolean isDefunct()

Returns true if all nodes in the group are running JE 7.5 or later, and therefore the GlobalCBVLSN is not maintained. Returns false for a mixed old/new version group (or when TEST_GLOBAL_CBVLSN is true).

getCBVLSN

public VLSN getCBVLSN()

Returns the global CBVLSN if it is not defunct. Returns a null if it is defunct.

For sake of old nodes, the global CBVLSN is computed as the minimum of CBVLSNs after discarding CBVLSNs that are obsolete. A CBVLSN is considered obsolete if it has not been updated within a configurable time interval relative to the time that the most recent CBVLSN was updated.

updateGroupInfo

public void updateGroupInfo(NameIdPair updateNameIdPair,
                            CleanerBarrierState barrierState)

Updates the cached group info for the node, avoiding a database read, if the global CBVLSN is not defunct. If it is defunct, does nothing.

Parameters:

updateNameIdPair - the node whose localCBVLSN must be updated.

barrierState - the new node syncup state

recalculate

public void recalculate(RepGroupImpl groupInfo)

Recalculates the GlobalVLSN when it is not defunct. When it is defunct, does nothing.

For sake of old nodes, the globalCBVLSN is computed as it was earlier: the minimum of CBVLSNs after discarding CBVLSNs that are obsolete. A CBVLSN is considered obsolete if it has not been updated within a configurable time interval relative to the time that the most recent CBVLSN was updated.

Note that the read of GroupInfo is not protected, and that groupInfo could be changing. That's okay, because we guarantee that none of the local CBVLSNs can be LT globalCBVLSN. If a local CBVLSN is written, and we miss it, it only means that this recalculation of global CBVLSN is too pessimistic -- it's too low.

Secondary nodes do not appear in the RepGroupDB, but the feeder has local CBVLSN values for them which are part of this calculation. Secondary nodes and new nodes have their VLSN ranges protected by the mechanism which tries to ensure that files which may be needed by active feeders are not candidates for deletion.

getRestoreResponseVLSN

public VLSN getRestoreResponseVLSN(VLSNRange range)

Returns a VLSN appropriate for the RestoreResponse.cbvlsn field when the GlobalCBVLSN is not defunct. When it is defunct, returns a null VLSN.

When sending a RestoreResponse to a (potentially) old node, we supply a VLSN that will cause selection of reasonably current nodes as a server (feeder) for the network restore. The return value is the VLSN range end of this node (the master) minus a value that will allow up-to-date servers to qualify (NETWORKBACKUP_MAX_LAG). Old nodes will reject servers whose VLSN range does not cover this VLSN.

In JE 7.4 and earlier, the "group CBVLSN" was used for this value. This was incorrect, because it was the lower bound for lagging replicas and reserved files. In JE 7.5 we improve on this by sending the value described above, when older nodes must be supported.

When GlobalCBLVSN is defunct, a null VLSN is returned because the RestoreResponse.cbvlsn field is not used at all. See the updated 'Algorithm' in NetworkRestore.