Filter-Based LSM Index Acceleration

Motivation [Back to TOC]

Traditional relational databases usually employ conventional index structures such as B+ trees due to their low read latency. However, such traditional index structures use in-place writes to perform updates, resulting in costly random writes to disk. Today’s emerging applications often involve insert-intensive workloads for which the cost of random writes prohibits efficient ingestion of data. Consequently, popular NoSQL systems such as Cassandra, HBase, LevelDB, BigTable, etc. have adopted Log-Structured Merge (LSM) Trees as their storage structure. LSM-trees avoids the cost of random writes by batching updates into a component of the index that resides in main memory – an in-memory component. When the space occupancy of the in-memory component exceeds a specified threshold, its entries are flushed to disk forming a new component – a disk component. As disk components accumulate on disk, they are periodically merged together subject to a merge policy that decides when and what to merge. The benefit of the LSM-trees comes at the cost of possibly sacrificing read efficiency, but, it has been shown in previous studies that these inefficiencies can be mostly mitigated.

AsterixDB has also embraced LSM-trees, not just by using them as primary indexes, but also by using the same LSM-ification technique for all of its secondary index structures. In particular, AsterixDB adopted a generic framework for converting a class of indexes (that includes conventional B+ trees, R trees, and inverted indexes) into LSM-based secondary indexes, allowing higher data ingestion rates. In fact, for certain index structures, our results have shown that using an LSM-based version of an index can be made to significantly outperform its conventional counterpart for both ingestion and query speed (an example of such an index being the R-tree for spatial data).

Since an LSM-based index naturally partitions data into multiple disk components, it is possible, when answering certain queries, to exploit partitioning to only access some components and safely filter out the remaining components, thus reducing query times. For instance, referring to our TinySocial example, suppose a user always retrieves tweets from the TweetMessages dataset based on the send-time field (e.g., tweets posted in the last 24 hours). Since there is not a secondary index on the send-time field, the only available option for AsterixDB would be to scan the whole TweetMessages dataset and then apply the predicate as a post-processing step. However, if disk components of the primary index were tagged with the minimum and maximum timestamp values of the objects they contain, we could utilize the tagged information to directly access the primary index and prune components that do not match the query predicate. Thus, we could save substantial cost by avoiding scanning the whole dataset and only access the relevant components. We simply call such tagging information that are associated with components, filters. (Note that even if there were a secondary index on send-time field, using filters could save substantial cost by avoiding accessing the secondary index, followed by probing the primary index for every fetched entry.) Moreover, the same filtering technique can also be used with any secondary LSM index (e.g., an LSM R-tree), in case the query contains multiple predicates (e.g., spatial and temporal predicates), to obtain similar pruning power.

Filters in AsterixDB [Back to TOC]

We have added support for LSM-based filters to all of AsterixDB’s index types. To enable the use of filters, the user must specify the filter’s key when creating a dataset, as shown below:

Creating a Dataset with a Filter

    create dataset Tweets(TweetType) primary key tweetid with filter on send-time;

Filters can be created on any totally ordered datatype (i.e., any field that can be indexed using a B+ -tree), such as integers, doubles, floats, UUIDs, datetimes, etc.

When a dataset with a filter is created, the name of the filter’s key field is persisted in the Metadata.Dataset dataset (which is the metadata dataset that stores the details of each dataset in an AsterixDB instance) so that DML operations against the dataset can recognize the existence of filters and can update them or utilize them accordingly. Creating a dataset with a filter in AsterixDB implies that the primary and all secondary indexes of that dataset will maintain filters on their disk components. Once a filtered dataset is created, the user can use the dataset normally (just like any other dataset). AsterixDB will automatically maintain the filters and will leverage them to efficiently answer queries whenever possible (i.e., when a query has predicates on the filter’s key).

Filters and Merge Policies [Back to TOC]

The AsterixDB default merge policy, the prefix merge policy, relies on component sizes and the number of components to decide which components to merge. This merge policy has proven to provide excellent performance for both ingestion and queries. However, when evaluating our filtering solution with the prefix policy, we observed a behavior that can reduce filter effectiveness. In particular, we noticed that under the prefix merge policy, the disk components of a secondary index tend to be constantly merged into a single component. This is because the prefix policy relies on a single size parameter for all of the indexes of a dataset. This parameter is typically chosen based on the sizes of the disk components of the primary index, which tend to be much larger than the sizes of the secondary indexes’ disk components. This difference caused the prefix merge policy to behave similarly to the constant merge policy (i.e., relatively poorly) when applied to secondary indexes in the sense that the secondary indexes are constantly merged into a single disk component. Consequently, the effectiveness of filters on secondary indexes was greatly reduced under the prefix-merge policy, but they were still effective when probing the primary index. Based on this behavior, we developed a new merge policy, an improved version of the prefix policy, called the correlated-prefix policy. The basic idea of this policy is that it delegates the decision of merging the disk components of all the indexes in a dataset to the primary index. When the policy decides that the primary index needs to be merged (using the same decision criteria as for the prefix policy), then it will issue successive merge requests to the I/O scheduler on behalf of all other indexes associated with the same dataset. The end result is that secondary indexes will always have the same number of disk components as their primary index under the correlated-prefix merge policy. This has improved query performance, since disk components of secondary indexes now have a much better chance of being pruned.