Automating memory management in a DBMS

I recently read two papers on automating memory management - one for DB2, one for Oracle. While the papers aren't new (~2005), they are definitely worth reading. AFAIK both made it into the products, I have no idea how that turned out but the outcome doesn't change my opinion that the papers were excellent. Disclaimer - I worked on query execution for Oracle at the time and made a few changes to the row sources that I maintained to support the work described by the Oracle paper.

Both describe methods to automate memory management with the goal of improving performance. In many DBMS memory management is difficult for several reasons:

• It isn't global and the DBA must estimate good values for the sizes of various caches and other large memory consumers: sort and hash for order by, aggregation, join and index create.
• It isn't bounded when allocations are specified per instance of sort or join rather than one limit on the memory used by all sorts or joins. Too many concurrent queries == OOM in this case.
• It is static. This is an issue for large caches like the buffer pool (block cache).

There was a marketing battle between Oracle and IBM in the mid-2000s over autonomic computing and these papers, along with getting them into products, is one outcome from that battle.

From the DB2 paper

A summary:

1. Differentiable functions were created to explain the query latency saved per page of memory for a variety of consumers. All functions had the same form (see section 3.2) with graphs that look like this. I think that time saved means wall clock time to account for CPU and IO overheads.
2. For a few caches (statement and buffer pool) online simulators were added to the DBMS to estimate changes to the hit rate if the cache were given more memory. 
3. Constrained optimization was used to determine the optimal allocation at any point in time. The constraint was the amount of memory available. I assume that Lagrange multipliers were used.
4. Feedback control was used to apply the desired memory configuration. One benefit from this is to avoid negative impacts from suddenly applying significant changes.
5. Decisions are revisited because workloads (types of queries, concurrency) changes.

Comments:

• At least for sort, the real curve for time vs memory can't be described by a differentiable function. See page 2 in the Oracle paper: there are three intervals where each interval can be explained by a function but the points where the intervals meet are a problem. I am not sure whether constrained optimization can handle that.
• At a higher level, the paper doesn't consider the steps in the time vs memory benefits for some row sources (this is a kind of a repeat of the previous comment).
• It wasn't clear whether row sources could give back memory. There is a risk from giving a long-running sort or hash a lot of memory. If there is no facility to get back some of that memory then memory allocation will be far from optimal until that query completes or is killed.

From the Oracle paper

While Oracle has automatic memory management for caches (plan cache, buffer pool) and queries (hash, sort, bitmap index operators) the paper is limited to memory management for queries (PGA in Oracle terminology).

Summary:

1. for each instance of a row source (a query is composed of multiple row sources, some row sources can use a lot of memory for sort and hash) the knees in the response time vs memory graph are estimated, see section 3.2 in the paper. These knees are a function of the row source and the data specific to a query (query A might be able to sort in-memory with 100M of RAM while query B might require 1G to remain in memory).
2. decisions are made about the amount of memory that each row source can used based on the information from the previous point
3. over time a row source might be able to get more memory and might be told to give back memory. Giving back memory might not be immediate, but should eventually be done.

Comments:

• the paper was vague about how point #2 was done. While section 4.2.2 lists 5 rules used to guide the decisions it wasn't clear there was a goal beyond making sure all queries have enough memory to run. In contrast the IBM paper was clear that constrained optimization was used and explains the function that was optimized.

Developer experience, what about the other stakeholders?

While developer experience gets a lot of press, there are four stakeholders when you provide a database service. So we can call these DX, UX, MX and OX for developer, user, management and operations experience:

• developers want the DBMS to stay out of their way. Schema is one example because waiting for a schema change gets in the way. Note that NoSQL databases have less schema rather than being schema-less because indexes are schema. I am curious whether less schema leads to a great developer experience in the long run for large scale projects given the risk of an unmanaged schema and poorly understood data. The developers who were able to move fast early in the project can create much tech debt for those who arrive years later.
  
• users of the services that depend on the database want great QoS - high uptime, low and predictable latency for queries, low chance of lost data. They just want to use your database-backed app without problems.
• management wants to minimize cost while getting great QoS.
• operations wants to be able to sleep when they are oncall (self-healing database, auto failover, etc). It helps if the database isn't in chaos mode during working hours as that gives them freedom to get their work done.

Review of DiffKV

This is a review of Differentiated Key-Value Storage Management for Balanced I/O Performance that was published in ATC 2021. This is a wonderful paper that resolves several of my concerns about key-value separation (here and here). The authors have experience with key-value separation via Titan which is part of TiDB.

The key idea in the paper is to use leveled compaction for keys and tiered compaction for values. Write-amplification is reduced by not using leveled compaction for values. Read-amplification is reduced, relative to classic key-value separation (see WiscKey), by using tiered compaction for values rather than logs. With classic key-value separation the worst case read-amp for a scan is the need to do a random read from the log for every qualifying key as that random read can include a block decompression and a storage IO. With DiffKV the use of tiered compaction for values provides some ordering to avoid that worst case.

While I enjoyed the paper, it didn't have performance results with compression enabled and I am curious about the impact of decompression overhead on point and range read latency.

A summary of the approach:

• small values are stored inline in the LSM tree
• large values are stored in logs, called vLogs. The paper explains a few improvements to make GC faster in this case. But mostly this is classic key-value separation.
• medium values use the vTree (tiered compaction for values)
• the user defines the two size limits that determines whether a value is small, medium or large

The vTree

Medium values are stored in the vTree which resembles an LSM that uses tiered compaction. A vTable is the unit of storage in the vTree. A vTable is 8MB by default and stores KV pairs in key order. A sorted group is a sequence vTables for which the keys don't overlap -- (vTable, sorted group) are similar to an (SST, sorted run) in RocksDB. A vTree has multiple levels where each level has one or more sorted groups and the levels have exponentially increasing size.

This is similar to tiered compaction for two reasons. First, there are multiple sorted groups per level. Second, when merges are done to move values from level N to N+1, the merge process only reads data from level N and then writes it to level N+1. It doesn't read and rewrite data already on level N+1. 

To reduce the number of times that values get rewritten the vTree has fewer levels than the LSM tree that stores keys. The smallest N levels in the key's LSM tree share the smallest vTree level. I assume the remaining levels in the key's LSM tree each have their own level in the vTree.

vTree GC

Keys in the LSM tree point into the vTree. When a value is moved between levels in the vTree the key's entry in the LSM tree must be updated to point to the new location in the vTree. DiffKV couples GC in the vTree with GC in the key's LSM tree to avoid extra writes. By coupling I mean that vTree GC is extra work added to compaction done on the key's LSM tree. When a key is moved from level N to level N+1 in the LSM tree then its value might be moved to the next level in the vTree.

By coupling it like this, DiffKV avoids the need to probe the index (key's LSM tree) to determine whether a value is live -- which is an extra overhead that occurs with classic key-value separation. It also avoids generating extra writes to change the point into the vTree that is stored with the key.

This above is called compaction-triggered merge and driven by compaction of the key's LSM tree. Another reason for doing vTree GC is called scan-triggered and triggered by scans that encounter regions of the vTree that need GC.

I am curious whether there is a need to also trigger GC based on vTables that have excessive space-amplification.

vTree rewrites

Leveled compaction has more write-amp than tiered because it rewrites previously written KV pairs. While vTree GC frequently avoids the need to do rewrites, there is one case where a rewrite is needed and I didn't learn enough about how DiffKV handles this.

There can be space wasted from vTables that have low utilization rates because most of the values in the vTable have been deleted. In this case something should be done to copy out (rewrite) the values to a new vTable. And when values are moved the key LSM tree must be updated to reference their new location. For workloads with updates and deletes this will be needed in the largest level of the vTree and might be needed for smaller levels.