Building Applications with MongoDB's Pluggable Storage Engines: Part 1

Jason Ma


This is the first in a two post series about MongoDB’s pluggable storage engines. This post discusses characteristics of MongoDB’s storage engines.


With users building increasingly complex data-driven apps, there is no longer a "one size fits all" database storage technology capable of powering every type of application built for the enterprise. Modern applications need to support a variety of workloads with different access patterns and price/performance profiles – from low latency, in-memory read and write applications, to real time analytics, to highly compressed "active" archives.

Through the use of pluggable storage engines, MongoDB can be extended with new capabilities, and configured for optimal use of specific hardware architectures. This approach significantly reduces developer and operational complexity compared to running multiple database technologies. Storage engines can be mixed in the same replica set or sharded cluster. Users can also leverage the same MongoDB query language, data model, scaling, security and operational tooling across different applications, each powered by different pluggable MongoDB storage engines.

**Figure 1:** Mix and match storage engines within a single MongoDB replica set

MongoDB 3.2 ships with four supported storage engines that can be optimized for specific workloads:

  • The default WiredTiger storage engine. For most applications, WiredTiger's granular concurrency control and native compression will provide the best all-around performance and storage efficiency.
  • The Encrypted storage engine, protecting highly sensitive data, without the performance or management overhead of separate file system encryption. The Encrypted storage engine is based upon WiredTiger and so throughout this whitepaper, statements regarding WiredTiger also apply to the Encrypted storage engine. This engine is part of MongoDB Enterprise Advanced.
  • The In-Memory storage engine for applications that have extremely strict SLAs for consistent and predictable low latency, while not requiring disk durability for the data. This engine is part of MongoDB Enterprise Advanced. The MMAPv1 engine, an improved version of the storage engine used in pre-3.x MongoDB releases. MMAPv1 was the default storage engine in MongoDB 3.0.

MongoDB allows users to mix and match multiple storage engines within a single MongoDB cluster. This flexibility provides a simple and reliable approach to support diverse workloads. Traditionally, multiple database technologies would need to be managed to meet these needs, with complex, custom integration code to move data between technologies, and to ensure consistent, secure access. With MongoDB’s flexible storage architecture, the database automatically manages the movement of data between storage engine technologies using native replication. This approach significantly reduces developer and operational complexity when compared to running multiple distinct database technologies.

**Table 1:** Comparing the MongoDB WiredTiger, In-Memory, Encrypted, and MMAPv1 storage engines

**WiredTiger Storage Engine**

MongoDB acquired WiredTiger in 2014, and with it the experts behind the WiredTiger storage engine: co-founders Keith Bostic (founder of Sleepycat Software) and Dr. Michael Cahill, and their colleagues. Bostic and Cahill were the original architects of Berkeley DB, the most widely-used embedded data management software in the world, and have decades of experience writing high performance storage engines. WiredTiger leverages modern hardware architectures and innovative software algorithms to provide industry-leading performance for the most demanding applications.

WiredTiger is ideal for wide range of operational applications and is therefore MongoDB’s default storage engine. It should be the starting point for all new applications, with the exception of cases where you need the specific capabilities of the In-Memory or Encrypted storage engines.

The key advantages of WiredTiger include:

Maximize Available Cache: WiredTiger maximizes use of available memory as cache to reduce I/O bottlenecks. There are two caches that are used: the WiredTiger cache and the filesystem cache. The WiredTiger cache stores uncompressed data and provides in-memory-like performance. The operating system’s filesystem cache stores compressed data. When data is not found in the WiredTiger cache, WiredTiger will look for the data in the filesystem cache.

**Figure 2:** WiredTiger Caches (WiredTiger Cache and FS Cache)

Data found in the filesystem cache first goes through a decompression process before moving to the WiredTiger cache.

The WiredTiger cache performs best when it holds as much of the working set as possible. However, it is also important to reserve memory for other processes that need it such as the operating system, including the filesystem cache. This also includes MongoDB itself, which as a whole will consume more memory than what is in active use by WiredTiger.

MongoDB defaults to a WiredTiger cache size of approximately 60% of RAM. The minimum amount to leave the filesystem cache is at 20% of available memory. Anything lower and the operating system may be constrained for resources.

High Throughput: WiredTiger uses “copy on write” — when a document is updated WiredTiger will make a new copy of the document and determine the latest version to return to the reader. This approach allows multiple clients to simultaneously modify different documents in a collection, resulting in higher concurrency and throughput. Optimum write performance is achieved when an application is utilizing a host with many cores (the more the better), and multiple threads are writing to different documents.

Reducing Storage Footprint and Improving Disk IOPs: WiredTiger uses compression algorithms to reduce the amount of data stored on disk. Not only is storage reduced, but IOPs performance is increased as fewer bits are read from or written to disk. Some types of files compress better than others. Text files are highly compressible, while binary data may not be as compressible since it may already be encoded and compressed. WiredTiger does incur additional CPU cycles when using compression, but users can configure compression schemes to optimize CPU overhead vs. compression ratio. Snappy, which is the default compression engine, provides good balance between high compression ratio with low CPU overhead. Zlib will achieve higher compression ratios, but incur additional CPU cycles.

Compression (Indexes and Journals): Indexes can be compressed in memory as well as on disk. WiredTiger utilizes prefix compression to compress the indexes, conserving RAM usage as well as freeing up storage IOPs. Journals are compressed by default with Snappy compression.

Multi-Core Scalability: As CPU manufacturers shrink to smaller lithographies and power consumption becomes more and more of an issue, processor trends have shifted to multi-core architectures in order to sustain the cadence of Moore’s law. WiredTiger was designed with modern, multi-core architectures in mind, and provides scalability across multi-core systems. Programming techniques such as hazard pointers, lock free algorithms, and fast latching minimize contention between threads. Threads can perform operations without blocking each other — resulting in less thread contention, better concurrency, and higher throughput.

Read Concern: WiredTiger allows users to specify a level of isolation for their reads. Read operations can return a view of data that has been accepted or committed to disk by a majority of the replica set. This provides a guarantee that applications only read data that will persist in the event of failure and won’t get rolled back when a new replica set member is promoted to primary.

For more information on migrating from MMAP/MMAPv1 to WiredTiger here is the documentation.

**Encrypted Storage Engine**

Data security is top of mind for many executives due to increased attacks as well as a series of data breaches in recent years that have negatively impacted several high profile brands. For example, in 2015, a major health insurer was a victim of a massive data breach in which criminals gained access to the Social Security numbers of more than 80 million people ­— resulting in an estimated cost of $100M. In the end, one of the critical vulnerabilities was the health insurer did not encrypt sensitive patient data stored at-rest.

Coupled with MongoDB’s extensive access control and auditing capabilities, encryption is a vital component in building applications that are compliant with standards such as HIPAA, FERPA, PCI, SOX, GLBA, ISO 27001, etc.

The Encrypted storage engine is based on WiredTiger, and thus is designed for operational efficiency and performance:

  • Document level concurrency control and compression
  • Support for Intel’s AES-NI equipped CPUs for acceleration of the encryption/decryption process
  • As documents are modified, only updated storage blocks need to be encrypted rather than the entire database

With the Encrypted storage engine, protection of data at-rest is an integral feature of the database. The raw database “plaintext” content is encrypted using an algorithm that takes a random encryption key as input and generates ciphertext that can only be decrypted with the proper key. The Encrypted Storage Engine supports a variety of encryption algorithms from the OpenSSL library. AES-256 in CBC mode is the default, while other options include AES-256 in GCM mode, as well as FIPS mode for FIPS-140-2 compliance.

Encryption is performed at the page level to provide optimal performance. Instead of having to encrypt/decrypt the entire file or database for each change, only the modified pages need to be encrypted or decrypted, resulting in less overhead and higher performance.

Additionally, the Encrypted Storage Engine provides safe and secure management of the encryption keys. Each encrypted node contains an internal database key that is used to encrypt or decrypt the data files. The internal database key is wrapped with an external master key, which must be provided to the node for it to initialize. To ensure that keys are never written or paged to disk in unencrypted form, MongoDB uses operating system protection mechanisms, such as VirtualLock and mlock, to lock the process’ virtual memory space into memory.

There are two primary ways to manage the master key: through an integration with a third party key management appliance via the Key Management Interoperability Protocol (KMIP) or local key management via a keyfile.

Most regulatory requirements mandate that the encryption keys be rotated and replaced with a new key at least once annually. MongoDB can achieve key rotation without incurring downtime by performing rolling restarts of the replica set. When using a KMIP appliance, the database files themselves do not need to be re-encrypted, thereby avoiding the significant performance overhead imposed by key rotation in other databases. Only the master key is rotated, and the internal database keystore is re-encrypted.

It is recommended to use a KMIP appliance with the Encrypted storage engine.

**In-Memory Storage Engine**

In modern applications, different subsets of application data have different latency and durability requirements. The In-Memory storage engine option is created for applications that have extremely strict SLAs even at 99th percentiles. The In-Memory engine will keep all of the data in memory, and will not write anything to disk. Data always has to be populated on start-up, and nothing can be assumed to be present on restart, including application data and system data (i.e users, permissions, index definitions, oplog, etc). All data must fit into the specified in-memory cache size.

The In-Memory storage engine combines the predictable latency benefits of an “in memory cache” with the rich query and analytical capabilities of MongoDB. It has the advantage of using the exact same APIs as any other MongoDB server so your applications do not need special code to interact with the cache, such as handling cache invalidation as data is updated. In addition, a mongod that's configured with the In-Memory storage engine can be part of a replica set, and thus can have another node in the same replica set backed by fast persistent storage.

The In-Memory engine is currently supported on MongoDB 3.2.6+. For performance metrics on the In-Memory storage engine view the MongoDB Pluggable Storage Engine white paper.

For applications requiring predictable latencies, the In-Memory engine is the recommended storage engine as it provides low latency while also minimizing tail latencies resulting in high performance and a consistent user experience.

Some of the key benefits of the In-Memory engine:

  • Predictable and consistent latency for applications that want to minimize latency spikes
  • Applications can combine separate caching and database layers into a single layer— all accessed and managed with the same APIs, operational tools, and security controls
  • Data redundancy with use of a WiredTiger secondary node in a replica set

**MMAPv1 Storage Engine**

The MMAPv1 engine is an improved version of the storage engine used in pre 3.x MongoDB releases. It utilizes collection level concurrency and memory mapped files to access the underlying data storage. Memory management is delegated to the operating system. This prevents compression of collection data, though journals are compressed with Snappy.

In the second part of this blog series, we will discuss how to select which storage engine to use.

Learn more about MongoDB’s pluggable storage engines. Read the whitepaper.

Pluggable Storage Engine Architecture

About the author - Jason Ma

Jason Ma is a Principal Product Marketing Manager based in Palo Alto, and has extensive experience in technology hardware and software. He previously worked for SanDisk in Corporate Strategy doing M&A and investments, and as a Product Manager on the Infiniflash All-Flash JBOF. Before SanDisk, he worked as a HW engineer at Intel and Boeing. Jason has a BSEE from UC San Diego, MSEE from the University of Southern California, and an MBA from UC Berkeley.