In Search of Database Nirvana

Data Structure—Key Support, Clustering, Partitioning

To handle all these different types of workloads, a query engine must first and foremost determine what kind of workload it is processing. Suppose that it is a single-row access. A single-row access could mean scanning all the rows in a very large table, if the structure does not have keyed access or any mechanism to reduce the scan. The query engine would need to know the key structure for the table to assess if the predicate(s) provided cover the entire key or just part of the key. If the predicate(s) cover the entire unique key, the engine knows this is a single-row access and the storage engine supporting direct keyed access can retrieve it very fast.

Statistics

Statistics are necessary when query engines are trying to generate query plans or understand whether a workload is operational or analytical. In the single-row-access scenario described earlier, if the predicate(s) used in the query only cover some of the columns in the key, the engine must figure out whether the predicate(s) cover the leading columns of the key, or any of the key columns. Let us assume that leading columns of the key have equality predicates specified on them. Then, the query engine needs to know how many rows would qualify, and how the data that it needs to access is spread across the nodes. Based on the partitioning scheme—that is, how data is spread across nodes and disks within those nodes—the query engine would need to determine whether it should generate a serial plan or a parallel plan, or whether it can rely on the storage engine to very efficiently determine that and access and retrieve just the right number of rows. For this, it needs some idea as to how many rows will qualify.

The only way for the query engine to know the number of rows that will qualify, so as to generate an efficient query plan, is to gather statistics on the data ahead of time to determine the cardinality of the data that would qualify. If multiple key columns are involved, most likely the cardinality of the combination of these columns is much smaller than the product of their individual cardinalities. So the query engine must have multicolumn statistics for key columns. Various statistics could be gathered. But at the least it needs to know the unique entry counts, and the lowest and highest, or second lowest and second highest, values for the column(s).

Skew is another factor to take into account. Skew becomes relevant when data is spread across a large number of nodes and there is a chance that a large amount of data could end up being processed by just a few nodes, overwhelming those nodes and affecting all of the workloads running on the cluster (given that most would need those nodes to run), whereas other nodes are waiting on these few nodes to finish executing the query. If the only types of workloads the query engine has to handle are OLTP or operational ones, the chances are it does not need to process large amounts of data and therefore does not need to worry about skew in the data, other than at the data partitioning layer, which can be controlled via the choice of a good partitioning key. But if it’s also processing BI and analytics workloads, skew could become an important factor. Skew also depends on the amount of parallelism being utilized to execute a query.

For situations in which skew is a factor, the database cannot completely rely on the typical equal-width histograms that most databases tend to collect. In equal-width histograms, statistics are collected with the range of values divided into equal intervals, based on the lowest and highest values found and the unique entry count calculated. However, if there is a skew, it is difficult to know which value has a skew because it would fall into a specific interval that has many other values in its range. So, the query engine has to either collect some more information to understand skew or use equal-height histograms.

Equal height histograms have the same number of rows in each interval. So if there is a skewed value, it will probably span a larger number of intervals. Of course, determining the right interval row size and therefore number of intervals, the adjustments needed to highlight skewed values versus nonskewed values (where not all intervals might end up having the same size) while minimizing the number of intervals without losing skew information is not easy to do. In fact, these histograms are a lot more difficult to compute and lead to a number of operational challenges. Generally, sampling is needed in order to collect these statistics fast, because the data must be sorted in order to put them into these interval buckets. You need to devise strategies to incrementally update these statistics and when to update them. These come with their own challenges.

Predicates on Nonleading Key Columns or Nonkey Columns

Things begin getting really tricky when the predicates are not on the leading columns of the key but are nonetheless on some of the columns of the key. What could make this more complex is an IN list against these columns with OR predicates, or even NOT IN conditions. A capability called Multidimensional Access Method (or MDAM) provides efficient access capabilities when leading key column values are not known. In this case, the multicolumn cardinality of leading column(s) with no predicates needs to be known in order to determine if such a method will be faster in accessing the data than a full table scan. If there are intermediate key columns with no predicates, their cardinalities are essential, as well. So, multikey column considerations are almost a must if these are not operational queries with efficient keys designed for their access.

Then, there are predicates on nonkey columns. The cardinality of these is relevant because it provides an idea as to the reduction in size of the resulting number of rows that need to be processed at upper layers of the query—such as joins and aggregates.

All of the above keyed and nonkeyed access cardinalities help determine join strategies and degree of parallelism.

If the storage engine is a columnar storage engine, the kind of compression used (dictionary, run length, and so on) becomes important because it affects scan performance. Also, the sequence in which these predicates should be evaluated becomes important in that case because you want to reduce as many rows as early as possible, so you want to begin with predicates on columns that give you the largest reduction first. Here too, clustered access versus a full scan versus efficient mechanisms to reduce scans of column values—which might be provided by the storage engine—are relevant. As are statistics.

Indexes and Materialized Views

Then, there is the entire area of indexing. What kinds of indexes are supported by the storage engine or created by the query engine on top of the storage engine? Indexes offer alternate access paths to the data that could be more efficient. There are indexes designed for index-only scans to avoid accessing the base table by having all relevant columns in the index.

There are also materialized views. Materialized views are relevant for more complex workloads for which you want to prematerialize joins or aggregates for efficient access. This is highly complex because now you need to figure out if the query can actually be serviced by a materialized view. This is called materialized view query rewrite.

Some databases call indexes and materialized views by different names, such as projections, but ultimately the goal is the same—to determine what the available alternate access paths are for efficient keyed or clustered access to avoid large, full-table scans.

Of course, as soon as you add indexes, a database now needs to maintain them in parallel. Otherwise, the total response time will increase by the number of indexes it must maintain on an update. It has to provide transactional support for indexes to remain consistent with the base tables. There might be considerations such as colocation of the index with the base table. The database must handle unique constraints. One example in BI and analytics environments (as well as some other scenarios) is that bulk loads might require an efficient mechanism to update the index and ensure that it is consistent.

Indexes are used more for operational workloads and much less so for BI and analytical workloads. On the other hand, materialized views, which are materialized joins and/or aggregations of data in the base table, and similar to indexes in providing quick access, are primarily used for BI and analytical workloads. The increasing need to support operational dashboards might be changing that somewhat. If materialized view maintenance needs to be synchronous with updates, they too can be a large burden on updates or bulk loads. If materialized views are maintained asynchronously, the impact is not as severe, assuming that audit logs or versioning can be used to refresh them. Some databases support user-defined materialized views to provide more flexibility to the user and not burden operational updates. The query engine should be able to automatically rewrite queries to take advantage of any of these materialized views when feasible.

Storage engines also use other techniques like Bloom filters and hash tables to speed access. The query engine needs to be aware of all the alternative access paths made available by the storage engine to get at the data. It also needs to know how to exploit them or implement them itself in order to deliver high performance for operational and analytical workloads.

Degree of Parallelism

All right, so now we know how we are going to scan a particular table, we have an estimate of rows that will be returned by the storage engine from these scans, and we understand how the data is spread across partitions. We can now consider both serial and parallel execution strategies, and balance the potentially faster response time of parallel strategies against the overhead of parallelism.

Yes, parallelism does not come for free. You need to involve more processes across multiple nodes, and each process will consume memory, compete for resources in its node, and that node is subject to failure. You also must provide each process with the execution plan, for which they must then do some setup to execute. Finally, each process must forward results to a single node that then has to collate all the data.

All of this results in potentially more messaging between processes, increases skew potential, and so on.

The optimizer needs to weigh the cost of processing those rows by using a number of potential serial and parallel plans and assess which will be most efficient, given the aforementioned overhead considerations.

To offer really high concurrency for all workloads (including large EDW workloads that can have a very large number of concurrent queries being executed in seconds or subseconds), the optimizer needs to assess the degree of parallelism needed for each query. To execute a query most efficiently in terms of response time and resources used, the query engine should base each operation’s degree of parallelism on the cardinality of rows that operation needs to process. Scans that filter rows, joins, and aggregates can often lead to substantial reduction in data. It makes no sense to use, say, 100 nodes to execute an operation when 5 nodes are sufficient to do so. Not only that, as soon as the maximum degree of parallelism required by the query—based on the cardinality of the data it will process—is known, the query can be allocated to run on a segment, or subset of the nodes, in the cluster. If the cluster were divided into a number of equal segments, it can be very efficiently used by allocating queries to run in those segments, or a combination of segments, thereby dramatically increasing concurrency. This yields the twin benefits of using system resources very efficiently while gaining more resiliency by reducing the degree of parallelism. This is illustrated in Figure 1-2.

As the cluster is expanded and newer technology is used for the added nodes, with potentially more resource capacity than existing nodes on the cluster, this segmentation can help use that capacity more efficiently by allocating more queries to the newer segment.

Reducing the Search Space

The options discussed so far provide optimizers a large number of potentially good query plans. There are various technologies such as Cascades, used by NonStop SQL (and now part of Apache Trafodion) and Microsoft SQL Server, that are great for optimizers but have the disadvantage of having this very large search space of query plans to evaluate. For long-running queries, spending extra time to find a better plan by trawling through more of that search space can have dramatic payoffs. But for operational queries, the returns of finding a better plan diminish very fast, and compile-time spent looking for a better plan becomes an issue, because most operational queries need to be processed within seconds or even subseconds.

One way to address this compile-time issue for operational queries is to provide query plan caching. These techniques cannot be naive string matching mechanisms alone, even after literals or parameters have been excluded. Table definitions could change since the last time the plan was executed. A cached plan might need to be invalidated in those cases. Schema context for the table could change, not obvious in the query text. A plan handling skewed values could be substantially different from a plan on values that are not skewed. So, sophisticated query plan caching mechanisms are needed to reduce the time it takes to compile while avoiding a stale or inefficient plan. The query plan cache needs to be actively managed to remove least recently used plans from cache to accommodate frequently used ones.

The optimizer can be a cost-based optimizer, but it must be rules driven, with the ability to add heuristics and rules very efficiently and easily as the optimizer evolves to handle different workloads. For instance, it should be able to recognize patterns. A star join is not likely in an operational query. But for BI queries, it could detect such a join. If it does, it can use specialized indexes designed for that purpose, or it could decide to do a cross product of the dimension tables (something optimizers otherwise avoid), before doing a nested join to the fact table, instead of scanning the entire fact table and doing repeated hash joins against the dimension tables.

Join Type

That brings us to join types. For operational workloads, a database needs to support nested joins and probe cache for nested joins. A probe cache for nested joins is where the optimizer understands that access to the inner table will have enough repetition due to the unsorted nature of the rows coming from the outer table, so that caching those results would really help with the join.

For BI and analytics workloads, a merge or hybrid hash join would most likely be more efficient. A nested join can be useful for such workloads some of the times. However, nested join performance tends to degrade rapidly as the amount of data to be joined grows.

Because a wrong choice can have a severe impact on query performance, you need to add a premium to the cost and not choose a plan purely on cost. Meaning, if there is a nested join with a slightly lower cost than a hash join, you don’t want to choose it, because the downside risk of it being a bad choice is huge, whereas the upside might not be all that better. This is because cardinality estimations are just that: estimations. If you chose a nested join or serial plan and the number of rows qualifying at run time are equal to or lower than compile time estimations, then that would turn out to be a good plan. However, if the actual number of rows qualifying at run time is much higher than estimated, a nested or serial plan might not be just bad, it can be devastating. So, a large enough risk premium can be assigned to nested joins and serial plans, so that hash joins and parallel plans are favored, to avoid the risk of a very bad plan. This premium can be adjusted, because different workloads respond differently to costing, especially when considering the balance between operational queries and BI or analytics queries.

For BI and analytics queries, if the data being processed by a hash join or a sort is large, detecting memory pressure and overflowing gracefully to disk is important. Operational queries, however, generally don’t have to deal with large amounts of data to the point that this is an issue.

Data Flow and Access

The architecture for a query engine needs to handle large parallel data flows with complex operations for BI and analytics workloads as well as quick direct access for operational workloads.

For BI and analytics queries for which larger amounts of data are likely to be processed, the query execution architecture should be able to parallelize at multiple levels. The first level is partitioned parallelism, so that multiple processes for an operation such as join or aggregation are executed in parallel. Second is at the operator level, or operator parallelism. That is, scans, multiple joins, aggregations, and other operations being performed to execute the query should be running concurrently. The query should not be executing just one operation at a time, perhaps materializing the results on disk in between as MapReduce does.

All processes should be executing simultaneously with data flowing through these operations from scans to joins to other joins and aggregates. That brings us to the third kind of parallelism, which is pipeline parallelism. To allow one operator in a query plan (say, a join) to consume rows as they are produced by another operator (say, another join or a scan), a set of up and down interprocess message queues, or intraprocess memory queues, are needed to keep a constant data flow between these operators (see Figure 1-3).

But for OLTP and operational queries, this data flow architecture (Figure 1-4) can be a huge overhead. If you are accessing a single row, or just a few rows, you don’t need the queues and complex data flows. In such a case, you can have optimizations to reduce the path length and quickly just access and return the relevant row(s).

While you are optimizing for OLTP queries with fast paths, for BI and analytics queries you need to consider prefetching blocks of data, provided the storage engine supports this, while the query engine is busy processing the previous block of data. So the nature of processing varies widely for the kind of workloads the query engine is processing, and it must accommodate all of these variants. Figures Figure 1-5 through Figure 1-8 illustrate how these processing scenarios can vary from a single row or single partition access serial plan or parallel multiple direct partition access for an operational query, to multitiered parallel processing of BI and analytics queries to facilitate complex aggregations and joins.

Mixed Workload

One of the biggest challenges for HTAP is the ability to handle mixed workloads; that is, both OLTP queries and the BI and analytics queries running concurrently on the same cluster, nodes, disks, and tables. Workload management capabilities in the query engine can categorize queries by data source, user, role, and so on and allow users to prioritize workloads and allocate a higher percentage of CPU, memory, and I/O resources to certain workloads over others. Or, short OLTP workloads can be prioritized over BI and analytics workloads. Various levels of sophistication can be used to manage this at the query level.

However, storage engine optimization is required, as well. The storage engine should automatically reduce the priority of longer running queries and suspend execution on a query when a higher priority query needs to be serviced, and then go back to running the longer running query. This is called antistarvation, because you don’t want to starve out higher priority queries from running, or even same or lower priority queries from running, while a single query hogs all the resources. An alternate way to address this might be to direct update workloads to the primary partition for a specific row being updated and query workloads to its replicates if the storage engine can facilitate this without loss of consistency.

Streaming

More and more applications need incoming streams of data processed in real time, necessitating the application of functions, aggregations, and trigger actions across a stream of data, often time-series data, over row count or time-based windows. This is very different from processing statistical or user-defined functions, sophistical algorithms, aggregates, and even Online Analytical Processing (OLAP) window functions over data persisted in a table on disk or memory. Even though Jennifer Widom had proposed new SQL syntax to handle streams in 2008, there is no standard SQL syntax to process streaming data. Query engines have to be able to deal with this new data processing paradigm.

Feature Support

Last but not least is the list of features you need to support for operational and analytical workloads. These features range from referential integrity, stored procedures, triggers, various levels of transactional isolation and consistency, for operational workloads; to materialized views; fast/bulk Extract, Transform, Load (ETL) capabilities; and OLAP, time series, statistical, data mining, and other functions for BI and analytics workloads.

Features common to both types of workloads are many. Some of the capabilities a query engine needs to support are scalar and table mapping UDFs, inner, left, right, and full outer joins, un-nesting of subqueries, converting correlated subqueries to joins, predicate push down, sort avoidance strategies, constant folding, recursive union, and so on.

This is not close to an exhaustive list, but supporting all these capabilities for these different workloads takes a huge investment of resources.

Statistics

As was discussed earlier, statistics are needed to support any query workload in order to generate a good query plan or to even understand whether the workload is operational or analytical. So, when a new storage engine is added, you need to add support for statistics gathering. Does the storage engine have an efficient mechanism to sample rows? Does it have an easy way to query new data for incremental stats computations? Does it already gather count and other metrics that can be used for fast stats collection? Does it gather counts on records deleted, changed, or inserted, so that the query engine can determine when to schedule another update of its statistics? You need to answer a number of questions to set up the gathering of statistics to work efficiently with the storage engine.

Key Structure

For operational workloads, keyed access is important for subsecond response times. Single-row access requires access via a key. Ideally multicolumn keys are supported, because invariably transactional or fact tables have multicolumn keys. If multicolumn keys are not supported, the query engine needs to map the multiple primary key columns to the single storage engine key. Also, short operational queries often require clustered access to retrieve a small number of rows. Range access over clustered keys helps operational workloads meet Service Level Agreements (SLAs). The query engine needs to understand what keyed access options are available for the storage engine in order to appropriately exploit them for the most efficient access. It needs to optimize this access for each storage engine it supports.

Partitioning

How the storage engine partitions data across disks and nodes is also very important for the query engine to understand. Does it support hash and/or range partitioning, or a combination of these? How is this partitioning determined? Does the query engine need to salt data so that the load is balanced across partitions in order to avoid bottlenecks? If it does, how can it add a salt key, say as the leftmost column of the table key, and still avoid table scans? Does the storage engine handle repartitioning or rebalancing of partitions as the cluster is expanded or contracted, or does the query engine need to worry about doing that? This can get very complex because users might need full read and write access while this repartitioning of data is occurring. How the data is spread across partitions is very important if the query engine is going to have parallel processes working on data from these partitions. It needs to try to localize access and reduce repartitioning of data (also called shuffle), or sending data across nodes, as much as possible. Without understanding the key and partitioning structures of the storage engine, it is impossible to efficiently process data for a query. Operational queries might not often encounter some of these challenges of processing data in parallel. But how data is partitioned can influence whether fast operational queries can be serviced by access to a single partition, disk, or node, versus having to access multiple partitions to service a query with stringent service level objectives.

Data Type Support

The query engine needs to understand what data types are supported by the storage engine, and what constraints the storage engine enforces on those types, so that the query engine can enforce any constraints that the storage engine doesn’t. For example, in some cases, it is better for the query engine to store the data in its own encoded data type if the storage engine just supports string arrays. Or, if the query engine enforces CHECK constraints, again, can it subcontract that to the storage engine, or does it need to enforce them? Certainly, referential constraints enforcement is not something a storage engine will likely take on.

Does the storage engine provide full character set support so that UTF-8 support between the query and storage engine is sufficient for the query engine to support all character sets? What about collations?

Does the storage engine provide compression or encryption support, or will the query engine need to provide that support at its level? The query engine will most likely be the one issuing the table creation DDL statements and needs to know which options for block level and key compression, and other table structure options need to be exposed by it to the user. In some cases, a storage engine table might be mapped to a query engine’s view of that table.

Projection and Selection

Most storage engines support projection, wherein only the desired columns are returned by the storage engine. However, if that is not the case, the query engine must unpack the data and do the projection. Similarly, predicate evaluation provided by the storage engine needs to be understood. What predicates will the storage engine be able to evaluate (=, <, >, <>, between, LIKE)? Can it evaluate predicates on multiple columns, or multicolumn predicates ((a,b) > (5,1))? Can it handle an IN list? How long an IN list? Generally, can it handle a combination of ORs and ANDs, and to what level of complexity? Does it automatically determine the most efficient sequence in which to apply the predicates, such as evaluating predicates on a column that yields the highest reduction in cardinality first, especially if the storage engine is a column store? Can it handle predicates involving multiple columns of the same table; for instance, c1 > c2? Besides literals, can it handle string expressions (STR(c1, 3, 2)='10'), or date and time expressions? How complex can these expressions be? How does the storage engine handle default or missing values (NULLs)? The query engine needs to understand all of these aspects of a storage engine in order to determine what predicates it can push to the storage engine and which ones it needs to implement itself.

As memory prices have gone down, it has become possible to configure large amounts of memory per node. The promise of nonvolatile memory is around the corner. Not only are storage engines designing efficient structures in memory for access to data for both transactional and analytical workloads, but they are implementing techniques to make operations such as projection and selection more efficient. They are taking advantage of L1, L2, and L3 CPU cache to process data at even faster speeds than main memory, implementing vectorization to process sets of rows/column values at a time, and coming up with other innovative ways, such as minimizing or eliminating the cost of serialization/deserialization, to push data processing speeds to the ultimate while utilizing the hardware to the utmost. This moves the bottleneck from disk to main memory, and ultimately to the CPU. Efficient use of the CPU now becomes imperative. The query engine must be able to exploit these storage engine efficiencies and be frugal in the use of CPU itself.

Extensibility

Some storage engines can run user-defined code, such as coprocessors in HBase, or before and after triggers, in order to implement more functionality on the storage engine/server side to reduce the amount of message traffic and improve efficiencies.

For example, the optimizer might be smart enough to do eager aggregation if it can push that aggregation to the storage engine. This is especially effective if the data is clustered on the GROUP BY columns or if the number of groupings are expected to be small. If the storage engine does not support this, the query engine could use coprocessors to do the same. Complex predicate evaluations with expressions and functions, collocated joins and index maintenance, security enforcement, some ANSI Trigger actions, and even transactional support can be performed at this level.

Security Enforcement

Handling security is another point of contact between the query engine and the storage engine. The storage engine might have its own underlying security implementation. If it is well integrated with the SQL model, the query engine can use it. Otherwise, the query engine must administer privileges to schemas, tables, columns, and even rows in the case of fine-grained access control. It might need to integrate with, and even map, its authorization framework to that of the storage engine. If there are other considerations related to security, for example Hadoop security solutions such as Sentry or Ranger managing these objects, the query engine needs to integrate with that security framework. Depending on support for security logging and other Security Information and Event Management (SIEM) capabilities, the query engine must be able to integrate with those storage engine and platform capabilities as well, or provide its own. With HTAP, this is all made more complicated by the fact that there might be multiple storage engines that need to be supported.

Transaction Management

We can assume that a storage engine will provide some level of replication for high availability, backup and restore, and some level of multi–data center support. But transactional support can be a different thing altogether. There are the Atomic, Consistent, Isolated, Durable (ACID) and Basic Availability, Soft-state, Eventual consistency (BASE) models of transaction management. Depending on the support provided by the storage engine, the query engine might need to provide full multirow, multitable, and multistatement transaction support. It also needs to provide online backup, transactionally consistent recovery, often to a point-in-time, and multi–data center support with active-active synchronous updates as well as eventual consistency. It needs to be able to integrate this implementation with write-ahead logging and other mechanisms of the storage engine so as to take advantage of the underlying replication and other high-availability capabilities of the engine. These can be very different, as is already the case with Cassandra, HBase, or Hive, for example; each has its own transactional model.

Also, going back to extensibility of storage engine capabilities, such as coprocessors in HBase, the query engine needs to implement this transactional capability as close to and integrated with the storage engine as possible to be efficient and distributed. Otherwise, it will end up with a more generalized implementation, but not a scalable and efficient one.

Metadata Support

The query engine needs to add metadata support for the storage engine tables being supported. There are potential mappings (such as catalog, name, location, as well as data type mappings); extended options specific to the storage engine (such as compression options); supporting multiple column families—for example, mapping data partition options to underlying storage engine structures (e.g., HBase regions); and so on. When someone changes the storage engine table externally, how is this metadata updated, or how is it marked as invalid until someone fixes the discrepancy?

In fact, if there is access to the storage engine external to the query engine, there is a whole can of worms to deal with, such as how can transactional consistency and data integrity be guaranteed? What should the query engine do if the data in a column is not consistent with the data type it is supposed to have, or contains invalid data, perhaps because it has been updated by some other means? Also, do you allow secondary indexes, views, constraints, materialized views, and so on to be created on such a table? Not only does metadata support need to be provided for this support, but the issue of inconsistency potentially caused by external access has to be addressed.

There can be differences in metadata support needed for operational workloads versus BI and analytic workloads. For example, operational features such as referential integrity, triggers, and stored procedures need metadata support that is not needed for BI workloads, whereas metadata support for materialized views is not often needed for operational workloads.

Performance, Scale, and Concurrency Considerations

There are various performance, scaling, and concurrency considerations for each different storage engine with which a query engine needs to integrate in order to support HTAP workloads. For instance, is there a bulk-load mechanism available and what are its implications on transactional consistency and availability of the table during such loads? Are rowset inserts available so that large amounts of inserts can be performed efficiently? Are bulk-extract capabilities available, with sets of rows being returned in each buffer, instead of just a row at a time interface? Are faster scan options available, such as snapshot scans? Is the ability to request prefetching of data in large blocks available, when the query engine knows that a large amount of data is expected to be scanned, perhaps for an extract or for large complex queries? How can the query engine parallelize updates and access to the storage engine? If the query engine supports executor processes to parallelize queries, how best should it access the storage engine files/partitions and regions? What levels of concurrent access can the storage engine support? How best can the query engine utilize what the storage engine supports and compensate for what it does not? This is tough to determine, and more difficult to implement.

Error Handling

Error handling, too, is different for each storage engine that a query engine might need to support for HTAP. What are the error handling mechanisms available in the storage engine and how are these errors logged? Does the storage engine log errors or does the query engine need to interpret each type of error and log it, as well as provide some meaningful error message and remediation guidance to the user?

Other Operational Aspects

Although error handling is certainly one operational aspect, there are other considerations as well. For example, in the case of HBase there are issues with compaction, or splitting. These have significant impact on performance and user workloads, let alone transactional workloads. How will these be managed by the query engine to provide the best customer experience?

Suffice it to say that supporting a storage engine is not a trivial task. Sure, you can take shortcuts; support, say, InputFormat or OutputFormat; or even instrument a UDF or some connector technology to make the query engine talk to the storage engine. But to really integrate with a storage engine and exploit it to its best capabilities and provide the most efficient parallel implementation, it takes a lot of work and effort. If you need to support multiple storage engines to reach the database nirvana of one query engine on top of multiple storage engines, know that your task is cut out for you.

Just as the industry came together to create standard interfaces to query engines such as Java Database Connectivity (JDBC) and Open Database Connectivity (ODBC), it would be nice if disparate storage engines could provide a more common interface to query engines. Query engines could quickly and with less effort identify the capabilities of the storage engine via such an interface, integrate with it, and exploit its full capabilities. Maybe the industry will reach a point of maturity at which this could happen. For now, the APIs for the various storage engines and their capabilities are very diverse.

Although some of these challenges look very similar to the challenges that data federation engines must deal with, the difference is that data federation engines are interfacing with query engines using the standard JDBC/ODBC interfaces; in other words, subcontracting the query out to another query engine. This level of abstraction is very different from the much deeper level of integration required between a query engine and a storage engine geared toward a much higher degree of efficiency and performance. Just imagine splitting a proprietary RDBMS such as Oracle into a query and storage engine. That level of integration is quite a different than a federation engine interfacing with an Oracle instance.

High Availability

Availability is an important metric for any database. You can get around 99.99 percent availability with HBase, for example. Four nines, as they are known, means that you can have about 52.56 minutes of unscheduled downtime per year. Many mission-critical applications strive for five nines, or 5.26 minutes of downtime per year. Of course, the cost of that increase in availability can be substantial. The question is this: what high-availability characteristics can the query and storage engine combined provide?

Typically, high availability is very difficult for databases to implement, with the need to address the following questions:

• Can you upgrade the underlying OS, Hadoop distribution, storage or query engine to a new version of software while the data is available not only for reads, but also for writes, with no downtime—in other words, support rolling upgrades?

• Can you redistribute the data across new nodes or disks that have been added to the cluster, or consolidate them on to fewer nodes or disks, completely online with no downtime? Related to this is the ability to repartition the data for whatever reason. Can you do that online? Online could mean just read access or both read and write access to the data during the operation.

• Can you make online DDL changes to the database, such as changing the data type of a column, with no impact on reads and writes? Addition and deletion of nonkey columns have been relatively easy for databases to do.

• Can you create and drop secondary indexes online?

• Is there support for online backups, both full and incremental?

Which of the preceding capabilities your applications need will depend on the mix of operational and analytical workloads you have as well as how important high availability is for those workloads.

Security

Security implementations between operational and analytical workloads can be very different. For operational workloads, generally security is managed at the application level. The application interfaces with the user and manages all access to the database. On the other hand, BI and analytics workloads could have end users working with reporting and analytical tools to directly access the database. In such cases, there is a possibility that authorization is pushed to the database and the query and storage engine need to manage that security.

Integration into SIEM systems could be pertinent for analytical workloads, as well. But certainly many operational workloads require a higher degree of security controls and visibility.

Manageability

One of the most important aspects of a database is the ability to manage it and its workloads. As you can see in Figure 1-9, manageability entails a long list of things, and perhaps can only be partially implemented.

Given the mixed nature of HTAP workloads, some management tasks become increasingly challenging, especially workload management. Accounting for an OLTP or operational workload is generally done toward a “transaction” or interaction. The idea is to assess service level objectives in terms of transactions per second. Because the latency of such transactions can be so small, tracking performance at a per-statement, or even transaction level, would be very expensive. Thus, you need to accumulate the statistics at some interval and average it out. On the BI and analytics side, you generally do this assessment at a report or query level. These queries are generally longer, and capturing metrics on them and monitoring and managing based on query-level statistics could work just fine.

Managing mixed workloads to SLAs can be very challenging, based on priority and/or resource allocation. The ad hoc nature of analytics does not blend well with operational workloads that require consistent subsecond response times, even at peak loads.

If the query engine is integrating with different storage engines, even getting all the metrics for a query can be challenging. Although the query engine can track some metrics, such as time taken by the storage engine to service a request, it might not have visibility to what resources are being consumed by the underlying storage engine. How can you collect the vital CPU, memory, and I/O metrics, as well as numerous other metrics like queue length, memory swaps, and so on when that information is split between the query engine and the storage engine and there is nothing to tie that information together? For example, if you want to get the breakdown of the query resource usage by operation (for every step in the query plan—scan, join, aggregation, etc.) and by table accessed, especially when these operations are executed in parallel and the tables are partitioned, this becomes a very difficult job, unless the implementation has put in instrumentation and hooks to gather that end-to-end information. How do you find out why you have a skew or bottleneck, for example?

This does not purport to be an exhaustive list of challenges related to enterprise-level operational capabilities that are needed if these varied type of workloads are going to run on a single platform. There is the integration with YARN or Mesos, for example, which would take a lot more deliberation about how to manage resources for different application workloads across a cluster. Hopefully, though, it gives you a good sense for the challenges at hand.

Capabilities of the Query Engine

The considerations to assess the capabilities of a query engine of course depends on what kinds of workloads you want to run. But because this report is about supporting a mixed HTAP workload, here are some of the considerations that are relevant:

Integration Between the Query and Storage Engines

The considerations to assess the integration between the query and storage engines begins with understanding what capabilities you need a storage engine to provide. Then, you need to assess how well the query engine exploits and expands on those capabilities, and how well it integrates with those storage engines. Here are certain points to consider, which will help you determine not only if they are supported, but also at what level are they being supported: by the query engine or storage engine, or a combination of the two:

Data Model Support

Here are the considerations to assess the data model support:

Enterprise-Caliber Capabilities

Security was covered earlier, but here are the other considerations to assess enterprise-caliber capabilities: