Replication
Replication Options
Replicated directories are a fundamental requirement for delivering a resilient enterprise deployment.
Digi-CA™ has various configuration options for creating a replicated directory. In previous releases, replication was discussed in terms of a master server and some number of slave servers. A master accepted directory updates from other clients, and a slave only accepted updates from a (single) master. The replication structure was rigidly defined and any particular database could only fulfill a single role, either master or slave.
Digi-CA™ now supports a wide variety of replication topologies, these terms have been deprecated in favor of provider and consumer: A provider replicates directory updates to consumers; consumers receive replication updates from providers. Unlike the rigidly defined master/slave relationships, provider/consumer roles are quite fluid: replication updates received in a consumer can be further propagated by that consumer to other servers, so a consumer can also act simultaneously as a provider. Also, a consumer need not be an actual LDAP server; it may be just an LDAP client.
The following sections will describe the replication technology and discuss the various replication options that are available.
LDAP Sync Replication
The LDAP Sync Replication engine, syncrepl for short, is a consumer-side replication engine that enables the consumer LDAP server to maintain a shadow copy of a DIT fragment. A syncrepl engine resides at the consumer and executes as one of the slapd(8) threads. It creates and maintains a consumer replica by connecting to the replication provider to perform the initial DIT content load followed either by periodic content polling or by timely updates upon content changes.
Syncrepl uses the LDAP Content Synchronization protocol (or LDAP Sync for short) as the replica synchronization protocol. LDAP Sync provides a stateful replication which supports both pull-based and push-based synchronization and does not mandate the use of a history store. In pull-based replication the consumer periodically polls the provider for updates. In push-based replication the consumer listens for updates that are sent by the provider in realtime. Since the protocol does not require a history store, the provider does not need to maintain any log of updates it has received (Note that the syncrepl engine is extensible and additional replication protocols may be supported in the future.).
Syncrepl keeps track of the status of the replication content by maintaining and exchanging synchronization cookies. Because the syncrepl consumer and provider maintain their content status, the consumer can poll the provider content to perform incremental synchronization by asking for the entries required to make the consumer replica up-to-date with the provider content. Syncrepl also enables convenient management of replicas by maintaining replica status. The consumer replica can be constructed from a consumer-side or a provider-side backup at any synchronization status. Syncrepl can automatically resynchronize the consumer replica up-to-date with the current provider content.
Syncrepl supports both pull-based and push-based synchronization. In its basic refreshOnly synchronization mode, the provider uses pull-based synchronization where the consumer servers need not be tracked and no history information is maintained. The information required for the provider to process periodic polling requests is contained in the synchronization cookie of the request itself. To optimize the pull-based synchronization, syncrepl utilizes the present phase of the LDAP Sync protocol as well as its delete phase, instead of falling back on frequent full reloads. To further optimize the pull-based synchronization, the provider can maintain a per-scope session log as a history store. In its refreshAndPersist mode of synchronization, the provider uses a push-based synchronization. The provider keeps track of the consumer servers that have requested a persistent search and sends them necessary updates as the provider replication content gets modified.
With syncrepl, a consumer server can create a replica without changing the provider's configurations and without restarting the provider server, if the consumer server has appropriate access privileges for the DIT fragment to be replicated. The consumer server can stop the replication also without the need for provider-side changes and restart.
Syncrepl supports partial, sparse, and fractional replications. The shadow DIT fragment is defined by a general search criteria consisting of base, scope, filter, and attribute list. The replica content is also subject to the access privileges of the bind identity of the syncrepl replication connection.
The LDAP Content Synchronization Protocol
The LDAP Sync protocol allows a client to maintain a synchronized copy of a DIT fragment. The LDAP Sync operation is defined as a set of controls and other protocol elements which extend the LDAP search operation. This section introduces the LDAP Content Sync protocol only briefly. For more information, refer to RFC4533.
The LDAP Sync protocol supports both polling and listening for changes by defining two respective synchronization operations: refreshOnly and refreshAndPersist. Polling is implemented by the refreshOnly operation. The consumer polls the provider using an LDAP Search request with an LDAP Sync control attached. The consumer copy is synchronized to the provider copy at the time of polling using the information returned in the search. The provider finishes the search operation by returning SearchResultDone at the end of the search operation as in the normal search. Listening is implemented by the refreshAndPersist operation. As the name implies, it begins with a search, like refreshOnly. Instead of finishing the search after returning all entries currently matching the search criteria, the synchronization search remains persistent in the provider. Subsequent updates to the synchronization content in the provider cause additional entry updates to be sent to the consumer.
The refreshOnly operation and the refresh stage of the refreshAndPersist operation can be performed with a present phase or a delete phase.
In the present phase, the provider sends the consumer the entries updated within the search scope since the last synchronization. The provider sends all requested attributes, be they changed or not, of the updated entries. For each unchanged entry which remains in the scope, the provider sends a present message consisting only of the name of the entry and the synchronization control representing state present. The present message does not contain any attributes of the entry. After the consumer receives all update and present entries, it can reliably determine the new consumer copy by adding the entries added to the provider, by replacing the entries modified at the provider, and by deleting entries in the consumer copy which have not been updated nor specified as being present at the provider.
The transmission of the updated entries in the delete phase is the same as in the present phase. The provider sends all the requested attributes of the entries updated within the search scope since the last synchronization to the consumer. In the delete phase, however, the provider sends a delete message for each entry deleted from the search scope, instead of sending present messages. The delete message consists only of the name of the entry and the synchronization control representing state delete. The new consumer copy can be determined by adding, modifying, and removing entries according to the synchronization control attached to the SearchResultEntry message.
In the case that the LDAP Sync provider maintains a history store and can determine which entries are scoped out of the consumer copy since the last synchronization time, the provider can use the delete phase. If the provider does not maintain any history store, cannot determine the scoped-out entries from the history store, or the history store does not cover the outdated synchronization state of the consumer, the provider should use the present phase. The use of the present phase is much more efficient than a full content reload in terms of the synchronization traffic. To reduce the synchronization traffic further, the LDAP Sync protocol also provides several optimizations such as the transmission of the normalized entryUUIDs and the transmission of multiple entryUUIDs in a single syncIdSet message.
At the end of the refreshOnly synchronization, the provider sends a synchronization cookie to the consumer as a state indicator of the consumer copy after the synchronization is completed. The consumer will present the received cookie when it requests the next incremental synchronization to the provider.
When refreshAndPersist synchronization is used, the provider sends a synchronization cookie at the end of the refresh stage by sending a Sync Info message with refreshDone=TRUE. It also sends a synchronization cookie by attaching it to SearchResultEntry messages generated in the persist stage of the synchronization search. During the persist stage, the provider can also send a Sync Info message containing the synchronization cookie at any time the provider wants to update the consumer-side state indicator.
In the LDAP Sync protocol, entries are uniquely identified by the entryUUID attribute value. It can function as a reliable identifier of the entry. The DN of the entry, on the other hand, can be changed over time and hence cannot be considered as the reliable identifier. The entryUUID is attached to each SearchResultEntry or SearchResultReference as a part of the synchronization control.
Syncrepl Details
The syncrepl engine utilizes both the refreshOnly and the refreshAndPersist operations of the LDAP Sync protocol. If a syncrepl specification is included in a database definition, slapd(8) launches a syncrepl engine as a slapd(8) thread and schedules its execution. If the refreshOnly operation is specified, the syncrepl engine will be rescheduled at the interval time after a synchronization operation is completed. If the refreshAndPersist operation is specified, the engine will remain active and process the persistent synchronization messages from the provider.
The syncrepl engine utilizes both the present phase and the delete phase of the refresh synchronization. It is possible to configure a session log in the provider which stores the entryUUIDs of a finite number of entries deleted from a database. Multiple replicas share the same session log. The syncrepl engine uses the delete phase if the session log is present and the state of the consumer server is recent enough that no session log entries are truncated after the last synchronization of the client. The syncrepl engine uses the present phase if no session log is configured for the replication content or if the consumer replica is too outdated to be covered by the session log. The current design of the session log store is memory based, so the information contained in the session log is not persistent over multiple provider invocations. It is not currently supported to access the session log store by using LDAP operations. It is also not currently supported to impose access control to the session log.
As a further optimization, even in the case the synchronization search is not associated with any session log, no entries will be transmitted to the consumer server when there has been no update in the replication context.
The syncrepl engine, which is a consumer-side replication engine, can work with any backends. The LDAP Sync provider can be configured as an overlay on any backend, but works best with the back-bdb or back-hdb backend.
The LDAP Sync provider maintains a contextCSN for each database as the current synchronization state indicator of the provider content. It is the largest entryCSN in the provider context such that no transactions for an entry having smaller entryCSN value remains outstanding. The contextCSN could not just be set to the largest issued entryCSN because entryCSN is obtained before a transaction starts and transactions are not committed in the issue order.
The provider stores the contextCSN of a context in the contextCSN attribute of the context suffix entry. The attribute is not written to the database after every update operation though; instead it is maintained primarily in memory. At database start time the provider reads the last saved contextCSN into memory and uses the in-memory copy exclusively thereafter. By default, changes to the contextCSN as a result of database updates will not be written to the database until the server is cleanly shut down. A checkpoint facility exists to cause the contextCSN to be written out more frequently if desired.
Note that at startup time, if the provider is unable to read a contextCSN from the suffix entry, it will scan the entire database to determine the value, and this scan may take quite a long time on a large database. When a contextCSN value is read, the database will still be scanned for any entryCSN values greater than it, to make sure the contextCSN value truly reflects the greatest committed entryCSN in the database. On databases which support inequality indexing, setting an eq index on the entryCSN attribute and configuring contextCSN checkpoints will greatly speed up this scanning step.
If no contextCSN can be determined by reading and scanning the database, a new value will be generated. Also, if scanning the database yielded a greater entryCSN than was previously recorded in the suffix entry's contextCSN attribute, a checkpoint will be immediately written with the new value.
The consumer also stores its replica state, which is the provider's contextCSN received as a synchronization cookie, in the contextCSN attribute of the suffix entry. The replica state maintained by a consumer server is used as the synchronization state indicator when it performs subsequent incremental synchronization with the provider server. It is also used as a provider-side synchronization state indicator when it functions as a secondary provider server in a cascading replication configuration. Since the consumer and provider state information are maintained in the same location within their respective databases, any consumer can be promoted to a provider (and vice versa) without any special actions.
Because a general search filter can be used in the syncrepl specification, some entries in the context may be omitted from the synchronization content. The syncrepl engine creates a glue entry to fill in the holes in the replica context if any part of the replica content is subordinate to the holes. The glue entries will not be returned in the search result unless ManageDsaIT control is provided.
Also as a consequence of the search filter used in the syncrepl specification, it is possible for a modification to remove an entry from the replication scope even though the entry has not been deleted on the provider. Logically the entry must be deleted on the consumer but in refreshOnly mode the provider cannot detect and propagate this change without the use of the session log on the provider.