The Adyen way of engineering: we design for 20x
At Adyen we take privacy and data security very seriously, and it is at the heart of everything that we build. As a global payment platform, we handle a lot of Personally Identifiable Information or PII, which is any information that identifies a person. We use this information to drive many of our powerful tools such as anti-money laundering, identity verification (commonly referred to as KYC), fraud detection and more.
Being part of the greater payment industry, which moves around trillions of dollars every year, we are built on layers of trust. Our customers trust Adyen and we need to make sure that we are always in control of data security.
We are headquartered in the EU and with many of our customers doing business in the EU (with EU-based users); GDPR compliance is crucial. We also hold a European Banking license, which brings with it several degrees of compliance and regulatory requirements over and above other types of business that handles Personal Data.
Being a payment service provider means that we are a visible and valuable target to cyber criminals. We work continuously to make sure that we’re handling our data with care. We are constantly evaluating ourselves, minimizing risks at every step of the way with the ultimate goal of protecting our customers and their user’s data.
In this blog post, we’ll look at the challenges of securely managing data, and how we built tools to scale secure data management.
Having large quantities of high quality data allows our systems to be accurate, feature-rich and powerful, but with great power comes great responsibility. That responsibility translates to two major problems — scalably handling all of the data (you can’t have different teams managing data differently), and handling it securely.
Addressing this responsibility starts with having a threat model. A threat model is a representation of an application (Adyen platform) and all potential security considerations and implications (data, architecture, malicious actors, etc). Adyen manages an extensive threat model.
Our threat model includes entities trying to either disrupt Adyen’s ability to provide service, or entities attempting to extract data from Adyen, such as by compromising a database. A compromised database could mean real impact to real people, as well as serious consequences to our business.
We include how compromised data could be used into our threat model:
Keeping the data safe is not enough — it has to scale. What works for a handful of transactions should work for a billion transactions without missing a beat. The system should be able to generate reports, detect fraud, handle payments, provide real-time monitoring, and keep powering all the other products in our system without any down-time.
We also have to have a system that can handle heavy reading and writing. A system tuned to handle a write-heavy operation, such as making a payment, might not work for a read-heavy operation such as report generation.
Lastly, we have a “scale 20x” standard when designing and building a system. If a potential solution can’t scale 20x, we won’t pursue it.
Along with the Right to privacy, jurisprudence and legislation in Europe and across the world is starting to identify a human right - the “right to be forgotten” - in law to allow "any person living a life of rectitude ... that right to happiness which includes a freedom from unnecessary attacks on his character, social standing or reputation." to quote some relevant case law. As part of our obligation to comply with the GDPR, we need to be able to technically support mechanisms that respect an individual’s right to be forgotten.
With the challenges of security, performance, scalability and compliance requirements established, we need to address how we store PII.
Everyone knows that storing passwords in plain text is a really bad idea. But what about PII? The damage that can be caused by leaked PII is just as bad as a stolen password, and potentially much worse — think identity theft.
Best practice password storage dictates using one-way cryptographic hashes with per password unique salt. This makes “stored” passwords difficult to compromise even with large a corpus of potential passwords. But this one-way approach typically does not work for PII data which needs to be passed back in the clear for display, reporting or matching processes.
Alternatively, we could encrypt it. Use a strong encryption algorithm, keep your encryption key safe, and the data is practically indecipherable. This is a very good strategy, and it scales very well. However it does not solve all our problems — what if we want to explore data to support business goals or generate reports? Reports need to contain representative data which can be used to visualize trends without compromising PII. Providing a report with a bunch of encrypted data is a bad idea for a couple of different reasons:
There are also issues with getting additional features you might need, like:
In order to overcome these problems and more, we use a different strategy called tokenization. Instead of returning cipher-texts to the customer or our downstream systems, we assign identifiers to the data and return that instead. These identifiers point to metadata that allows us to track various properties about the token itself as well as the data behind the token. Choosing a good identifier format and generation algorithm ensures that we can also scale in an environment which has multiple active tokenization services that operate concurrently in a high-availability striping or load-balancing structure.
Token generation at Adyen is implemented in a fairly straight-forward way:
Tokenization service data flow
PII data (this could be a name, email address, credit card details, etc) coming in from the left passes through our Tokenization system, which encrypts the data with a key from the HSM-backed vault key manager. The encrypted data is stored inside the Data Vault, and assigned a UUID that is unique for the context of the owner of the data, and the UUID becomes the token. The token then replaces the PII elements in the data, and stored in the Tokenized Data Lake, to be used by downstream processes.
When a Merchant-Facing consumer, such as a report builder, requires this data, it is read from the Tokenized Data Lake, and the generated report is detokenized using the Detokenization service, and the final report with actual data is generated.
Our internal services also require data, but they can now use the Tokenized data instead of the raw data, thus keeping the PII data safe.
This approach has several advantages, described below:
The UUIDs do not represent, in themselves, any information about what the underlying data is, which means that these tokens can be freely shared in our platform without risk of attacks such as enumeration, rainbow tables, collision or reversing, thereby creating an opaque reference to the data.
“Wait, you said they were context-free?”
While a token’s value is context-free, a token itself has a context which identifies for what purpose it was generated. Tokens generated for Merchant A will be different from tokens generated for Merchant B for the same data. This ensures that any compromise at Merchant A does not impact the tokens and data held for customers of Merchant A at other merchants.
Our payment systems that handle the firehose of data are distributed, thus we cannot have a single point of token generation, which cannot scale up for the continuously increasing volume of tokens we need to generate. If one tokenization instance becomes unavailable, we need to still keep the platform running — including tokenization — while we fix it. The only way to be fault tolerant while scaling up is to be able to keep the system consistent while being distributed, and using UUIDs is the first step towards that. UUIDs have a one-in-17 billion chance of collision, and are generated using hardware entropy.
Merchants can request Adyen to remove data held for a customer by providing relevant context information for the customer. As the process of reversing a token is controlled by the system that generates it, we can easily erase the tokens for the context on our platform, removing the possibility of the information being retrieved successfully.
PII should not be held longer than is necessary, but deleting data is sometimes hard — even finding all copies can be tough in large businesses. Tokenization services help, because the data is typically centralized in a vault. And since the data is stored in the vault under a rotating protection key, reversing is only possible for as long as the protection key is still active. Removing an older protection key by maintaining a retention policy on the key effectively makes the encryption irreversible.
Anyone who’s worked on a search system knows that the system works by breaking up given text into lexemes which are then indexed into searchable parts (we know that this is overly simplified but you should get the idea). Exact match search is easily achievable with the tokenization solution: we can directly match tokens.
Fuzzy search and wildcard search is not trivial with a tokenization-based system. Since there is no link between the data and its token, searching for Pasport or Passpor* can’t find Passport, which raises questions around how to index and search data fuzzily or provide auto-complete and query-correction reliably.
Adyen handles payments around the world — and our customers are also global. A person queuing to buy a coffee in San Francisco could also be buying a new dress from a fancy French online boutique at the same time. Our system needs to have a consistent view of this data internally to ensure that fraud risks, customer recognition and all the other things that people expect today can be managed effectively at speed across geographic scale. Generation of UUIDs needs to somehow be consistent globally, while also being random.
We design our systems keeping all parts of our platform in mind — development, security, infrastructure and the business. Tokens based on UUIDs might solve the issue of distributed generation, but the underlying storage needs to scale elegantly. Our choice of database must make sense across several metrics for us:
We are currently working on solving a set of interesting and hard problems, including:
If solving problems like these sounds interesting to you, reach out to us!
By submitting this form, you acknowledge that you have reviewed the terms of our Privacy Statement and consent to the use of data in accordance therewith.