Analyzing the Impact of GDPR on Storage Systems

03/09/2019
by   Vinay Banakar, et al.
0

The recently introduced General Data Protection Regulation (GDPR) is forcing several companies to make significant changes to their systems to achieve compliance. Motivated by the finding that more than 30 related to storage, we investigate the impact of GDPR compliance on storage systems. We illustrate the challenges of retrofitting existing systems into compliance by modifying Redis to be GDPR-compliant. We show that despite needing to introduce a small set of new features, a strict real-time compliance (eg logging every user request synchronously) lowers Redis' throughput by 95 Our work reveals how GDPR allows compliance to be a spectrum, and what its implications are for system designers. We discuss the technical challenges that need to be solved before strict compliance can be efficiently achieved.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

06/28/2019

Analyzing GDPR Compliance Through the Lens of Privacy Policy

With the arrival of the European Union's General Data Protection Regulat...
06/29/2018

Complying with Data Handling Requirements in Cloud Storage Systems

In past years, cloud storage systems saw an enormous rise in usage. Howe...
09/12/2019

A Note On Compliance Relations And Fixed Points

We study compliance relations between behavioural contracts in a syntax ...
08/22/2018

Are we there yet? Understanding the challenges faced in complying with the General Data Protection Regulation (GDPR)

The EU General Data Protection Regulation (GDPR), enforced from 25th May...
11/12/2018

Compliance in Real Time Multiset Rewriting Models

The notion of compliance in Multiset Rewriting Models (MSR) has been int...
06/07/2019

Lightweight Parallel Foundations: a model-compliant communication layer

We present the Lightweight Parallel Foundations (LPF), an interoperable ...
10/01/2021

Continuous Compliance using Calculated Event Log Layers

Compliance has traditionally been a reactive activity, where directives ...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

“In law, nothing is certain but the expense.”

Samuel Butler

Privacy and protection of personal data (or more aptly, the lack thereof) has become a topic of concern for the modern society. The gravity of personal data breaches is evident not only in their frequency (1300 in 2017 alone [19]) but also their scale (the Equifax breach [14] compromised the financial information of 145 million consumers), and scope (the Cambridge Analytica scandal [20] harvested personal data to influence the U.K. Brexit referendum and the 2016 U.S. Presidential elections). In response to this alarming trend, the European Union (EU) adopted a comprehensive privacy regulation called the General Data Protection Regulation (GDPR) [18].

GDPR defines the privacy of personal data as a fundamental right of all the European people, and accordingly regulates the entire lifecycle of personal data. Thus, any company dealing with EU people’s personal data is legally bound to comply with GDPR. While essential, achieving compliance is not trivial: Gartner estimates 

[13] that less than 50% of the companies affected by GDPR would likely be compliant by the end of 2018. This challenge is exacerbated for a vast majority of companies that rely on third-parties for infrastructure services, and hence, do not have control over the internals of such services. For example, a company running a webserver on Google Cloud would not be able to achieve compliance if the underlying cloud subsystems are violating the GDPR regulations.

Though GDPR governs the behavior of most of the infrastructure and operational components of an organization, its impact on the storage systems is potent: 31 of the 99 articles that make up GDPR directly pertain to storage systems. Motivated by this finding, we set out to investigate the impact of GDPR on storage systems. In particular, we ask the following questions: (i) What features should a storage system have to be GDPR-compliant? (ii) How does compliance affect the performance of different types of storage systems? (iii) What are the technical challenges in achieving strict compliance in an efficient manner?

By examining the GDPR articles, we identify a core set of (six) features that must be implemented in the storage layer to achieve compliance. We hypothesize that despite needing to implement a small set of new features, storage systems would experience a signficant performance impact. This stems from a key observation: GDPR’s goal of data protection by design and by default

sits at odd with the traditional system design goals (especially for storage systems) of optimizing for performance, cost, and reliability. For example, the regulation on identifying and notifying data breaches requires that a controller shall keep a record of all the interactions with personal data. From a storage system perspective, this turns every read operation into a read followed by a write.

To evaluate our hypothesis, we design and implement the changes required to make Redis, a widely used key-value store, GDPR-compliant. This not only illustrates the challenges of retrofitting existing systems into GDPR compliance but also quantifies the resulting performance overhead. Our benchmarking using YCSB demonstrates that the GDPR-compliant version experiences a 95% slowdown compared to the unmodified version.

We share several insights from our investigation. First, though GDPR is clear in its high-level goals, it is intentionally vague in its technical specifications. This allows GDPR compliance to be a continuum and not a fixed target. We define real-time compliance and eventual compliance to describe a system’s approach to completing GDPR tasks. Our experiments show the performance impact of this choice. For example, by storing the monitoring logs in a batch (say, once every second) as opposed to synchronously, Redis’ throughput improves by 6 while exposing it to the risk of losing one second worth of logs. Such tradeoffs present design choices for researchers and practitioners building GDPR-compliant systems. Second, some GDPR requirements sit at odds with the design principles and performance guarantees of storage systems. This could lead to storage systems offering differing levels of native support for GDPR compliance (with missing features expected to be handled by other infrastrucutre or policy components). Finally, we identify three key research challenges (namely, efficient deletion, efficient logging, and efficient metadata indexing) that must be solved to make strict compliance efficient.

2 Background on GDPR

GDPR [18] is laid out in 99 articles that describe its legal requirements, and 173 recitals that provide additional context and clarifications to these articles. GDPR is an expansive set of regulation that covers the entire lifecycle of personal data. As such, achieving compliance requires interfacing with infrastructure components (including compute, network, and storage systems) as well as operational components (processes, policies, and personnel). However, since our investigation primarily concerns with GDPR’s impact on storage systems, we focus on articles that describe the behavior of storage systems. These fall into two broad categories.

2.1 Rights of the Data Subject

There are 12 articles that codify the rights and freedoms of people (referred to as data subjects). Amongst these, four directly concern storage sytems.

The first one, Article 15: Right of access by the data subject allows any person whose personal data has been collected by a company to obtain detailed information about its usage including (i) the purposes of processing, (ii) the recipients to whom it has been disclosed, (iii) the period for which it will be stored, and (iv) its use in any automated decision-making. Thus, the storage system should not only be designed to store these metadata but also be organized to allow a timely access. Related to this is the Article 21: Right to object, which allows a person to object at any time to using their personal data for the purposes of marketing, scientific research, historical archiving, or profiling. This requires storage systems to know both whitelisted and blacklisted purposes associated with personal data at all times, and control access to it dynamically.

However, the most prominent one, Article 17: Right to be forgotten grants people the right to require the data controller to erase their personal data without undue delay. This right is broadly construed with a few exceptions made for law enforcement and public health. From a storage perspective, the article demands that the requested data be erased in a timely manner including all its replicas and backups. Finally, Article 20: Right to data portability states that people have the right to obtain all their personal information in a commonly used and machine-readable format as well as the right to have these trasmitted to another company directly. Thus, storage systems should have the capability to access and transmit all data belonging to a particular user in a timely fashion.

No. GDPR article Key requirement Storage feature 5.1 Purpose limitation Data must be collected and used for specific purposes Metadata indexing 5.1 Storage limitation Data should not be stored beyond its purpose Timely deletion 5.2 Accountability Controller must be able to demonstrate compliance All 13 Conditions for data collection Get user’s consent on how their data would be managed All 15 Right of access by users Provide users a timely access to all their data Metadata indexing 17 Right to be forgotten Find and delete groups of data Timely deletion 20 Right to data portability Transfer data to other controllers upon request Metadata indexing 21 Right to object Data should not be used for any objected reasons Metadata indexing 25 Protection by design and by default Safeguard and restrict access to data Access control, Encryption 30 Records of processing activity Store audit logs of all operations Monitoring 32 Security of data Implement appropriate data security measures Access control, Encryption 33, 34 Notify data breaches Share insights and audit trails from concerned systems Monitoring 46 Transfers subject to safeguards Control where the data resides Manage data location

Table 1: Key GDPR articles that significantly impact the design, interfacing, or performance of storage systems. The table maps the requirements of these articles into storage system features.

2.2 Responsibilities of the Data Controller

The next group of articles outline the responsibilities of entities that collect and use personal data (namely, the data controllers). There are 10 articles in this category that concern storage systems.

Three articles elucidate the high-level principles of data security and privacy that must be followed by all controllers. Article 24: Responsibility of the controller establishes that the ultimate responsibility for the security of all personal data lies with the controller that has collected it; Article 32: Security of processing requires the controller to implement risk-appropriate and state-of-the-art security measures including encryption and pseudonymization; and lastly, Article 25: Data protection by design and by default, specifies that all systems must be designed, configured, and administered with data protection as a primary goal.

There are several articles that set guidelines for the collection, processing, and transmission of personal data. The purpose limitation of Article 5: Processing of personal data manadates that personal data should only be collected for specific purposes and not be used for any other purposes. From a storage point, this translates to maintaining associated (purpose-)metadata that could be accessed and updated by systems that process personal data. Interestingly, Article 13 also ascertains that data subjects have the right to know the specific purposes for which their personal data would be used as well as the duration for which it will be stored. The latter requirement means that storage systems have to support time-to-live mechanisms in order to automatically erase the expired personal data.

Finally, while Article 30: Records of processing activities requires the controller to maintain logs of all activities concerning personal data, Article 33: Notification of personal data breach mandates them to notify the authorities and users within 72 hours of any personal data breaches. In conjunction with Accountability clause of Article 5 which puts the onus of proving compliance on the controller, these articles impose stringent requirements on storage systems: to monitor and maintain detailed logs of all control- and data-paths interactions. For instance, every read operation now has to be followed by a (logging-)write operation.

Table–1 summarizes these articles and translates their key requirements into specific storage features.

3 Designing for Compliance

Based on our analysis of GDPR, we identify six key features that a storage system must support to be GDPR-compliant. Then, we characterize how systems show variance in their support for these features.

3.1 Features of GDPR-Compliant Storage

Timely Deletion. Under GDPR, no personal data can be retained for an indefinite period of time. Therefore, the storage system should support mechanisms to associate time-to-live (TTL) counters for personal data, and then automatically erase them from all internal subsystems in a timely manner. GDPR allows TTL to be either a static time or a policy criterion that can be objectively evaluated.

Monitoring and Logging. In order to demonstrate compliance, the storage system needs an audit trail of both its internal actions and external interactions. Thus, in a strict sense, all operations whether in the data path (say, read or write) or control path (say, changes to metadata or access control) needs to be logged.

Indexing via Metadata. Storage systems should have interfaces to allow quick and efficient access to groups of data. For example, accessing all personal data that could be processed under a specific purpose, or exporting all data belonging to a user. Additionally, regulations also requires the ability to quickly retrieve and delete large amounts of data that match a criterion.

Access Control. As GDPR aims to limit access to personal data to only permitted entities, for established purposes, and for predefined duration of time, the storage system must support fine-grained and dynamic access control.

Encryption. GDPR mandates that personal data be encrypted both at rest and in transit. While pseudonymization may help reduce the scope and size of data needing encryption, it is still required and likely results in degradation of storage system performance.

Managing Data Location. Finally, GDPR restricts the geographical locations where personal data may be stored. This implies that storage systems should provide an ability to find and control the physical location of data at all times.

3.2 Degree of Compliance

Though GDPR is clear in its high-level goals, it is intentionally vague in its technical specifications. For example, GDPR mandates that no personal data can be stored indefinitely and must be deleted after its expiry time. However, it does not specify how soon after its expiry should the data be erased? Seconds, hours, or even days? GDPR is silent on this, only mentioning that the data should be deleted without an undue delay. What this means for system designers is that GDPR compliance need not be a fixed target, instead a spectrum. We capture this variance along two dimensions: response time and capability.

Real-time vs. Eventual Compliance. Real-time compliance is when a system completes the GDPR task (e.g., deleting expired data or responding to user queries) synchronously in real-time. Otherwise, we categorize it as eventually compliant. Given the steep penalties (up to 4% of global revenue or €20M, whichever is higher) for violating compliance, companies would do well to be in the strict end of the spectrum. However, as we demonstrate in §4, achieving real-time compliance results in signficantly high overhead unless the challenges outlined in §5.1 are solved. This problem is further exacerbated for organizations that operate at scale. For example, Google cloud platform informs [2] their users that for a deleted data to be completely removed from all their internal systems, it could take up to 6 months.

Full vs. Partial Compliance. Distinct from the response time, systems exhibit varying levels of feature granularities and capabilities. Such discrepancies arise because many GDPR requirements sit at odds with the design principles and performance guarantees of certain systems. For example, file systems do not implement indexing into files as a core operation since that feature is commonly supported via application software like grep. Similarly, many relational databases only partially and indirectly support TTL as that operation could be realized using user-defined triggers, albeit inefficiently. Thus, we define full compliance to be natively supporting all the GDPR features, and partial compliance as enabling feature support in conjunction with external infrastrucutre or policy components.

We use the term strict compliance to reflect that a system has achieved both full- and real-time compliance.

4 GDPR-Compliant Redis

Redis [3] is a prominent example of key-value stores, a class of storage where unstructured data (i.e., value) is stored in an associative array and indexed by unique keys. Our choice of Redis as the reference system is motivated by two reasons: (i) it is a modern storage system with an active open-source development, and (ii) key-value stores, in general, are not only widely deployed in Internet-scale systems  [10, 21, 16] but also an active area research  [6, 9, 15, 12, 22, 7, 17].

From amongst the features outlined in §3.1, Redis fully supports monitoring, metadata indexing, and managing data locations; partially supports timely deletion; offers no native support for access control and encryption. Below, we discuss our changes—some involving implementation while others simply concerning policy and configurations—towards making Redis, GDPR compliant.

Then, we evaluate the performance impact of our modifications to Redis (v4.0.11) using the Yahoo Cloud Serving Benchmark (YCSB)  [8]. We configure YCSB workloads to use 2M operations, and run them on a Dell Precision Tower 7810 with quad-core Intel Xeon 2.8GHz processor, 16 GB RAM, and 1.2TB Intel 750 SSD.

Figure 1: Performance overhead of GDPR-compliant Redis. Benchmarking with YCSB demonstrates that monitoring and encryption will each reduce Redis’ throughput to 30% of the original.

4.1 Monitoring and Logging

Redis offers several mechanisms to generate complete audit logs: a debugging command called MONITOR, configuring the server with slowlog option, and piggybacking on append-only-file (AOF). Our microbenchmarking revealed that since Redis anyway performs its journaling via AOF, the first two options result in more overhead than AOF. Also, MONITOR streams the logs over a network, thus requiring additional encryption. So, we selected the AOF approach. However, AOF records only those operations that modify the dataset. Thus, we had to update the AOF code to include all of Redis’ interactions. Our benchmarking shows that when we set AOF to fsync every operation to the disk synchronously, Redis’ throughput drops to 5% of its original. But interestingly, as Figure 1 shows, when we releaxed the fsync frequency to once every second, the performance improved by 6 i.e., throughput dropped only to 30% the original.

Key takeaway: Even fully supported features like logging can cause significant performance overheads. Interestingly, the overheads vary significantly based on how strictly the compliance is enforced.

4.2 Encryption

In lieu of natively extending Redis’ limited security model, we incorporate third-party modules for encryption. For data at rest, we use the Linux Unified Key Setup (LUKS) [1], and for data in transit, we set up transport layer security (TLS) using Stunnel  [4]. Figure 1 shows that Redis performs at a third of its original throughput when encryption is enabled. We observed that most of overhead was due to TLS: this was because the TLS proxies in our setup had reduced the average available network bandwidth from 44 Gbps to 4.9 Gbps, thereby affecting both latency and throughput of YCSB. While there are alternatives to the LUKS-TLS approach like key-level encryption, our investigation using the open-source Themis [5] cryptographic library showed similar performance overheads.

Key takeaway: Retrofitting new features, especially those that do not align with the core design philosophies, will result in excessive performance overheads.

4.3 Timely Deletion

While GDPR does not mandate a strict timeline for erasing personal data after a request has been issued, it does specify that such data be removed from everywhere without undue delays. Redis offers three groups of primitives to erase data: (i) DEL & UNLINK to remove one or more specified keys immediately, (ii) EXPIRE & EXPIREAT to delete a given key after a specified timeout period, and (iii) FLUSHDB & FLUSHALL to delete all the keys present in a given database or all existing databases respectively. The current mechanisms and policies of Redis present two hindrances.

The first issue concerns the lag between the time of request and time of actual removal. While most of the above commands erase the data proactively, taking a time proportional to the size of data being removed, EXPIRE*

commands take a passive approach. The only way to guarantee the removal of an expired key is for a client to proactively access it. In absense of this, Redis runs a lazy probabilistic algorithm: once every 100ms, it samples 20 random keys from the set of keys with expire flag set; if any of these twenty have expired, they are actively deleted; if less than 5 keys got deleted, then wait till the next iteration, else repeat the loop immediately. Thus, as percentage of keys with associated expire increases, the probability of their timely deletion decreases.

To quantify this delay in erasure, we populate Redis with keys, all of which have an associated expiry time. The time-to-live values are set up such that 20% of the keys will expire in short-term (5 minutes) and 80% in the long-term (5 days). Figure 2 then shows the time Redis took to completely erase the short-term keys once 5 minutes have elapsed. As expected, the time to erasure increases with the database size. For example, when there are 128k keys, clean up of expired keys (25k of them) took nearly 3 hours. To support a stricter compliance, we modify Redis to iterate through the entire list of keys with associated EXPIRE. Then, we re-run the same experiment to verify that all the expired keys are erased within sub-second latency for sizes of upto 1 million keys.

The second concern relates to the persistence of deleted data in subsystems beyond the main storage engine. For example, in Redis AOF persistence model, any deleted data persists in AOF until its compaction either via a policy-triggered or user-induced BGREWRITEAOF operation. Though Redis prevents any legitimate access to data that is already deleted, its decision to let these persist in various subsystems, purely for performance reasons, is antithetical to the GDPR goals besides exposing itself to side-channel attacks. A naive approach to guaranteeing an immediate removal of deleted personal data is to trigger AOF compaction every time a key gets deleted. However, since GDPR only mandates a reasonable time for clean up, it may be prudent to configure a periodic (say, hourly) AOF compaction, which in turn would guarantee that no deleted key persists beyond an hour boundary.

Key takeaway: Even when the system supports a GDPR feature, system designers should carefully analyze its internal data structures, algorithms, and configuration parameters to guage the degree of compliance.

Figure 2: The graph shows the delay in erasing the expired keys (20% of total keys in each case) beyond their TTL. In contrast, our GDPR-compliant Redis erases all the expired keys within sub-second latency.

5 Concluding Remarks

We analyze the impact of GDPR on storage systems. We find that achieving strict compliance efficiently is hard; a naive attempt at strict compliance results in significant slowdown. We modify Redis to be GDPR-compliant and measure the performance overhead of each modification. Below, we identify three key research challenges that must be addressed to achieve strict GDPR compliance efficiently.

5.1 Research Challenges

Efficient Logging. For strict compliance, every storage operation including reads must be synchronously written to persistent storage; persisting to solid state drives or hard drives results in significant performance degradation. New non-volatile memory technologies, such as Intel 3D Xpoint, can help reduce such overheads. Efficient auditing may also be achieved through the use of eidetic systems. For example, Arnold [11] is able to remember past state with only 8% overhead; adapting Arnold for GDPR remains a challenge.

Efficient Deletion. With all personal data possessing an expiry timestamp, we need data structures to efficiently find and delete (possibly large amounts of) data in a timely manner. Like timeseries databases, data can be indexed by their expiration time, then grouped and sorted by that index to speed up this process. Conceptually, storing a max heap of keys could also make deletes a lot faster. If a subroot of the tree is found to be expired, the entire subtree rooted at that subroot could be cut off.

Efficient Metadata Indexing. Several articles of GDPR require efficient access to groups of data based on certain attributes. For example, accessing all the keys that allow processing for a particular purpose while ignoring those that object to that purpose; or collating all the files of a particular user to be ported to a new controller. While traditional databases natively offer this ability via secondary indices, not all storage systems have efficient or configurable support for this capability.

5.2 Limitations and Importance

Given its preliminary nature, our work has several limitations. First, we investigate one particular system, Redis, from one category of storage namely, NoSQL stores. Expanding the scope to a broader range of storage systems say, relational databases, and file systems would increase the confidence of our findings. Next, our benchmarking is based on only one suite, the YCSB. Finally, it is likely that the performance of our GDPR-compliant Redis could be further improved with a deeper knowledge of Redis internals.

With the growing relevance of privacy regulations around the world, we expect this paper to trigger interesting conversations at HotStorage. This is one of the first efforts to systematically analyze the impact of GDPR on storge systems. We would be keen to engage the storage community in identifying and addressing the research challenges in this space.

References

  • [1] Cryptsetup and LUKS - open-source disk encryption. https://gitlab.com/cryptsetup/cryptsetup, Accessed Jan 2019.
  • [2] Data Deletion on Google Cloud Platform. https://cloud.google.com/security/deletion/, Accessed Jan 2019.
  • [3] Redis Data Store. https://redis.io, Accessed Jan 2019.
  • [4] Stunnel. https://www.stunnel.org, Accessed Jan 2019.
  • [5] Themis: Easy to use cryptographic framework for data protection. https://github.com/cossacklabs/themis, Accessed Jan 2019.
  • [6] Anirudh Badam, KyoungSoo Park, Vivek Pai, and Larry Peterson. HashCache: Cache Storage for the Next Billion. In USENIX NSDI, 2009.
  • [7] Oana Balmau, Diego Didona, Rachid Guerraoui, Willy Zwaenepoel, Huapeng Yuan, Aashray Arora, Karan Gupta, and Pavan Konka. TRIAD: Creating synergies between memory, disk and log in log structured key-value stores. In USENIX ATC, 2017.
  • [8] Brian Cooper, Adam Silberstein, Erwin Tam, Raghu Ramakrishnan, and Russell Sears. Benchmarking cloud serving systems with YCSB. In ACM SoCC, 2010.
  • [9] Biplob Debnath, Sudipta Sengupta, and Jin Li. SkimpyStash: RAM space skimpy key-value store on flash-based storage. In ACM SIGMOD, 2011.
  • [10] Giuseppe DeCandia, Deniz Hastorun, Madan Jampani, Gunavardhan Kakulapati, Avinash Lakshman, Alex Pilchin, Swaminathan Sivasubramanian, Peter Vosshall, and Werner Vogels. Dynamo: Amazon’s Highly Available Key-Value Store. In USENIX OSDI, 2007.
  • [11] David Devecsery, Michael Chow, Xianzheng Dou, Jason Flinn, and Peter M Chen. Eidetic Systems. In USENIX OSDI, 2014.
  • [12] Robert Escriva, Bernard Wong, and Emin Gün Sirer. HyperDex: A distributed, searchable key-value store. In ACM SIGCOMM, 2012.
  • [13] Amy Ann Forni and Rob van der Meulen. Organizations are unprepared for the 2018 European Data Protection Regulation. In Gartner, May 2017.
  • [14] Todd Haselton. Credit reporting firm equifax says data breach could potentially affect 143 million US consumers. In CNBC, Sep 7 2017.
  • [15] Hyeontaek Lim, Bin Fan, David Andersen, and Michael Kaminsky. SILT: A memory-efficient, high-performance key-value store. In ACM SOSP, 2011.
  • [16] Rajesh Nishtala, Hans Fugal, Steven Grimm, Marc Kwiatkowski, Herman Lee, Harry Li, Ryan McElroy, Mike Paleczny, Daniel Peek, Paul Saab, David Stafford, Tony Tung, and Venkateshwaran Venkataramani. Scaling Memcache at Facebook. In USENIX NSDI, 2013.
  • [17] Pandian Raju, Rohan Kadekodi, Vijay Chidambaram, and Ittai Abraham. PebblesDB: Building Key-Value Stores using Fragmented Log-Structured Merge Trees. In ACM SOSP, 2017.
  • [18] General Data Protection Regulation. Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 on the protection of natural persons with regard to the processing of personal data and on the free movement of such data, and repealing Directive 95/46. Official Journal of the European Union, 59(1-88), 2016.
  • [19] Victor Reklaitis. How the number of data breaches is soaring. In MarketWatch, May 25 2018.
  • [20] Olivia Solon. Facebook says Cambridge Analytica may have gained 37M more users’ data. In The Guardian, Apr 4 2018.
  • [21] Roshan Sumbaly, Jay Kreps, Lei Gao, Alex Feinberg, Chinmay Soman, and Sam Shah. Serving Large-scale Batch Computed Data with Project Voldemort. In USENIX FAST, 2012.
  • [22] Xingbo Wu, Yuehai Xu, Zili Shao, and Song Jiang. LSM-trie: an LSM-tree-based ultra-large key-value store for small data. In USENIX ATC, 2015.