|
Cache Control
Contents1. IntroductionOne of the strengths of mondrian's design is that you don't need to do any processing to populate special data structures before you start running OLAP queries. More than a few people have observed that this makes mondrian an excellent choice for 'real-time OLAP' -- running multi-dimensional queries on a database which is constantly changing. The problem is that mondrian's cache gets in the way. Usually the cache is a great help, because it ensures that mondrian only goes to the DBMS once for a given piece of data, but the cache becomes out of date if the underlying database is changing. This is solved with a set of APIs for cache control. Before I explain the API, let's understand how mondrian caches data. 2. How mondrian's cache worksMondrian's cache ensures that once a multidimensional cell -- say the Unit Sales of Beer in Texas in Q1, 1997 -- has been retrieved from the DBMS using an SQL query, it is retained in memory for subsequent MDX calculations. That cell may be used later during the execution of the same MDX query, and by future queries in the same session and in other sessions. The cache is a major factor ensuring that Mondrian is responsive for speed-of-thought analysis. The cache operates at a lower level of abstraction than access control. If the current role is only permitted to see only sales of Dairy products, and the query asks for all sales in 1997, then the request sent to Mondrian's cache will be for Dairy sales in 1997. This ensures that the cache can safely be shared among users which have different permissions. If the contents of the DBMS change while Mondrian is running, Mondrian's implementation must overcome some challenges. The end-user expects a speed-of-thought query response time yielding a more or less up-to-date view of the database. Response time necessitates a cache, but this cache will tend to become out of date as the database is modified. Mondrian cannot deduce when the database is being modified, so we introduce an API so that the container can tell Mondrian which parts of the cache are out of date. Mondrian's implementation must ensure that the changing database state does not yield inconsistent query results. Until now, control of the cache has been very crude: applications would typically call
to flush the cache which maps connect string URLs to in-memory datasets. The effect of this call is that a future connection will have to re-load metadata by parsing the schema XML file, and then load the data afresh. There are a few problems with this approach. Flushing all data and metadata is all appropriate if the contents of a schema XML file has changed, but we have thrown out the proverbial baby with the bath-water. If only the data has changed, we would like to use a cheaper operation. The final problem with the 3. CacheControl APIThe CacheControl API solves all of the problems described above. It provides fine-grained control over data in the cache, and the changes take place as soon as possible while retaining a consistent view of the data. When a connection uses the API to notify Mondrian that the database has changed, subsequent queries will see the new state of the database. Queries in other connections which are in progress when the notification is received will see the database state either before or after the notification, but in any case, will see a consistent view of the world. The cache control API uses the new concept of a cache region, an area of multidimensional space defined by one or more members. To flush the cache, you first define a cache region, then tell Mondrian to flush all cell values which relate to that region. To ensure consistency, Mondrian automatically flushes all rollups of those cells. 3.1. A simple exampleSuppose that a connection has executed a query:
and that this has populated the cache with the following segments:
Now suppose that the application knows that batch of rows from Oregon, Q2 have been updated in the fact table. The application notifies Mondrian of the fact by defining a cache region:
and flushing that region:
Now let's look at what segments are left in memory after the flush.
The effects are:
3.2. More about cell regionsThe previous example showed how to make a cell region consisting of a single member, and how to combine these regions into a two-dimensional region using a crossjoin. The CacheControl API supports several methods of creating regions:
The second overloading of
Recall that the cell cache is organized in terms of columns, not members. This makes member ranges difficult for mondrian to implement. A range such as "February 15th 2007 onwards" becomes
The region returned by
3.3. Merging and truncating segmentsThe current implementation does not actually remove the cells from memory. For instance, in segment YNS#1 in the example above, the cell (1997, USA, OR) is still in the segment, even though it will never be accessed. It doesn't seem worth the effort to rebuild the segment to save a little memory, but we may revisit this decision. In future, one possible strategy would be to remove a segment if more than a given percentage of its cells are unreachable. It might also be useful to be able to merge segments which have the same dimensionality, to reduce fragmentation if the cache is flushed repeatedly over slightly different bounds. There are some limitations on when this can be done, since predicates can only constrain one column: it would not be possible to merge the segments
3.4. Accessing the CacheControl API through olap4jThe CacheControl API is specific to Mondrian and is part of its private APIs. If your application was written at a higher abstraction level, say using olap4j, you will need to 'unwrap' the connection first and gain access to its private functions. With an olap4j connection, this can be done like so:
4. Other cache control topics4.1. Flushing the dimension cacheThe Cache Control API allows Mondrian integrators to modify the cache of dimension members. The main way that Mondrian caches dimensions in memory is via a cache of member children. That is to say, for a given member, the cache holds the list of all children of that member. If a dimension table row was inserted or deleted, or if its key attributes are updated, its parent's child list would need to be modified, and perhaps other ancestors too. For example, if a customer Zachary William is added in city Oakland, the children list of Oakland will need to be flushed. If Zachary is the first customer in Oakland, California's children list will need to be flushed to accommodate the new member Oakland. To enable control of the dimensions cache via the Cache Control API, one must first set The methods of the Cache Control API related to the dimensions cache are the following:
It is important to note that modifying the dimension cache will, as a side effect, also flush the corresponding regions within the cell cache. 4.2. Cache consistencyMondrian's cache implementation must solve several challenges in order to prevent inconsistent query results. Suppose, for example, a connection executes the query
It would be unacceptable if, due to updates to the underlying database, the query yielded a result where the total for [All gender] did not equal the sum of [Female] and [Male], such as
We cannot guarantee that the query result is absolutely up to date, but the query must represent the state of the database at some point in time. To do this, the implementation must ensure that both cache flush and cache population are atomic operations. First, Mondrian's implementation must provide atomic cache flush so that from the perspective of any clients of the cache. Suppose that while the above query is being executed, another connection issues a cache flush request. Since the flush request and query are simultaneous, it is acceptable for the query to return the state of the database before the flush request or after, but not a mixture of the two. The query needs to use two aggregates: one containing total sales, and another containing sales sliced by gender. To see a consistent view of the two aggregates, the implementation must ensure that from the perspective of the query, both aggregates are flushed simultaneously. The query evaluator will therefore either see both aggregates, or see none. Second, Mondrian must provide atomic cache population, so that the database is read consistently. Consider an example.
Atomic cache population is difficult to ensure if the database is being modified without Mondrian's knowledge. One solution, not currently implemented, would be for Mondrian to leverage the DBMS' support for read-consistent views of the data. Read-consistent views are expensive for the DBMS to implement (for example, in Oracle they yield the infamous 'Snapshot too old' error), so we would not want Mondrian to use these by default, on a database which is known not to be changing. Another solution might be to extend the Cache Control API so that the application can say 'this part of the database is currently undergoing modification'. This scenario has not even considered aggregate tables. We have assumed that aggregate tables do not exist, or if they do, they are updated in sync with the fact table. How to deal with aggregate tables which are maintained asynchronously is still an open question. 4.3. Metadata cache controlThe CacheControl API tidies up a raft of (mostly equivalent) methods which had grown up for controlling metadata (schema XML files loaded into memory). The methods
Author: Julian Hyde; last modified by Luc Boudreau, August 2011. |