Skip to content

kirilltitov/FDBSwift

Repository files navigation

FDBSwift v5 Swift: 5.5

Episode V: The Swift Awaitening

This is FoundationDB client for Swift. It's quite low-level, (almost) Foundationless and can into async/await (Swift-NIO not included and not required).

Installation

Obviously, you need to install FoundationDB first. Download it from official website. Next part is tricky because CFDB module (C bindings) won't link libfdb_c library on its own, and FoundationDB doesn't yet ship pkg-config during installation. Therefore you must install it yourself. Run

chmod +x ./scripts/install_pkgconfig.sh
./scripts/install_pkgconfig.sh

or copy scripts/libfdb.pc (choose your platform) to /usr/local/lib/pkgconfig/ on macOS or /usr/lib/pkgconfig/libfdb.pc on Linux.

Migration from v4 to v5

v5 is a huge update in terms of internals and API. The most important update is async/await adoption, of course. Because of that, Swift-NIO dependency has been dropped as redundant. Naturally, a lot of API just vanished as obsolete. However, the remaining API looks pretty much the same, except it's not using event loops and futures.

In v4 (and earlier versions) there was a blocking API. Naturally, barely anyone used it because, well, why would you. However, during v5 development it turned out that this API is a perfect foundation for adopting async/await, just add async in signature and you're ready to go. Like it's been waiting for it really. It's a good thing I've been maintaining this API along with NIO API. Good for me.

So now there are only two ways of using FDBSwift:

  1. Oneshot API with autocommit in AnyFDB (hence, FDB, the default impl), like get(key:), set(key:value:), atomic(_ op:key:value) etc.
  2. Transactional API which comes in two flavours:
    1. Manual transaction management with flat flow: fdb.begin(), some operations and transaction.commit() (basically you have full control over transaction and have to catch errors manually and process retry error yourself as well, see details and respective section below).
    2. Wrapped transactions: fdb.withTransaction { transaction in /* some operations */ } which manages retries for you, still you have to be aware of other errors.

One more not so minor change is that error case FDB.Error.transactionRetry doesn't have an AnyFDBTransaction associated value anymore. It was handy in v4 when you could begin a NIO transaction and not have a reference to it at ANY moment, including flatMapError/recover etc. (I personally had a lot of such cases). Now this problem is gone for good and this associated value isn't needed at all.

Usage

Root concepts

By default (and in the very core) this wrapper, as well as C API, operates with byte keys and values (not pointers, but Array<UInt8>). See Keys, tuples and subspaces section for more details.

Values are always bytes (typealias Bytes = [UInt8]) (or nil if key not found). Why not Data you may ask? I'd like to stay Foundationless for as long as I can (srsly, import half of the world just for Data object which is a fancy wrapper around NSData which is a fancy wrapper around [UInt8]?) (Hast thou forgot that you need to wrap all your Data objects with autoreleasepool or otherwise you get fancy memory leaks?) (except for Linux tho, yes), you can always convert bytes to Data with Data(bytes: myBytes) initializer (why would you want to do that? oh yeah, right, JSON... ok, but do it yourself please, extensions to the rescue).

Connection

// Default cluster file path depending on your OS
let fdb = FDB()

// OR
let fdb = FDB(clusterFile: "/usr/local/etc/foundationdb/fdb.cluster")

Optionally you may pass network stop timeout.

Keep in mind that at this point connection has not yet been established, it's automatically established on first actual database operation. If you would like to explicitly connect to database and catch possible errors, just call:

try fdb.connect()

Disconnect is automatic, on deinit. But you may also call disconnect() method directly. Be warned that if anything goes wrong during disconnect, you will get uncatchable fatal error. It's not that bad because disconnect should happen only once, when your application shuts down (and you shouldn't really care about fatal errors at that point). Also you very ought to ensure that FDB really disconnected before actual shutdown (trap SIGTERM signal and wait for disconnect to finish), otherwise you might experience undefined behaviour (I personally haven't really encountered that yet, but it's not a phantom menace; when you don't follow FoundationDB recommendations things get quite messy indeed).

Before you connected to FDB cluster you may also set network options:

try fdb.setOption(.TLSCertPath(path: "/opt/fdb/tls/chain.pem"))
try fdb.setOption(.TLSPassword(password: "changeme"))
try fdb.setOption(.buggifyEnable)

See FDB+NetworkOptions.swift file for complete set of network options.

Keys, tuples and subspaces

All keys are AnyFDBKey which is a protocol:

public protocol AnyFDBKey {
    func asFDBKey() -> Bytes
}

This protocol is adopted by String, StaticString, Tuple (NOT Tuple type from Swift), Subspace and Bytes (aka Array<UInt8>), so you may freely use any of these types, or adopt this protocol in your custom types.

Since you would probably like to have some kind of key namespacing in your application, you should stick to Subspace which is an extremely useful instrument for creating namespaces. Under the hood it utilizes the Tuple concept. You oughtn't really bother delving into it (in short: basically a discount MsgPack, a tricky binary protocol), just remember that currently subspaces accept String, Int, Float (aka Float32), Double, Bool, UUID, Tuple (hence FDBTuplePackable), FDB.Null (why would you do that?) and Bytes as arguments.

// dump subspace if you would like to see how it looks from the inside
let rootSubspace = FDB.Subspace("root")

// also check Subspace.swift for more details and usecases
let childSubspace = rootSubspace["child"]["subspace"]

// OR
let childSubspace = rootSubspace["child", "subspace"]

// Talking about tuples:
let tuple = FDB.Tuple(
    Bytes([0, 1, 2]),
    322,
    -322,
    FDB.Null(),
    "foo",
    FDB.Tuple("bar", 1337, "baz"),
    FDB.Tuple(),
    FDB.Null()
)
let packed: Bytes = tuple.pack()
let unpacked: FDB.Tuple = try FDB.Tuple(from: packed)
let tupleBytes: Bytes? = unpacked.tuple[0] as? Bytes
let tupleInt: Int? = unpacked.tuple[1] as? Int
// ...
let tupleEmptyTuple: FDB.Tuple? = unpacked.tuple[6] as? FDB.Tuple
let tupleNull: FDB.Null? = unpacked.tuple[7] as? FDB.Null
if tupleNull is FDB.Null || unpacked.tuple[7] is FDB.Null {}

// you get the idea

Setting values

Simple as that:

try await fdb.set(key: "somekey", value: someBytes)

// OR
try await fdb.set(key: Bytes([0, 1, 2, 3]), value: someBytes)

// OR
try await fdb.set(key: FDB.Tuple("foo", FDB.Null(), "bar", FDB.Tuple("baz", "sas"), "lul"), value: someBytes)

// OR
try await fdb.set(key: Subspace("foo", "bar"), value: someBytes)

Getting values

Value is always Bytes? (nil if key not found), you should unwrap it before use. Keys are, of course, still AnyFDBKeys.

let value = try await fdb.get(key: "someKey")

Range get (multi get)

Since in FoundationDB keys are lexicographically ordered over the underlying bytes, you can get all subspace values (or even from whole DB) by querying range from key somekey\x00 to key somekey\xFF (from byte 0 to byte 255). You shouldn't do it manually though, as Subspace object has a shortcut that does it for you.

Additionally, get(range:) (and its versions) method returns not Bytes, but a special box structure FDB.KeyValuesResult which holds an array of FDB.KeyValue structures and a flag indicating whether DB can provide more results (pagination, kinda):

public extension FDB {
    /// A holder for key-value pair
    public struct KeyValue {
        public let key: Bytes
        public let value: Bytes
    }
    
    /// A holder for key-value pairs result returned from range get
    public struct KeyValuesResult {
        /// Records returned from range get
        public let records: [FDB.KeyValue]

        /// Indicates whether there are more results in FDB
        public let hasMore: Bool
    }
}

If range call returned zero records, it would result in an empty FDB.KeyValuesResult struct (not nil).

let subspace = FDB.Subspace("root")
let range = subspace.range
/*
  these three calls are completely equal (can't really come up with case when you need second form,
  but whatever, I've seen worse whims)
*/
let result: FDB.KeyValuesResult = try await fdb.get(range: range)
let result: FDB.KeyValuesResult = try await fdb.get(begin: range.begin, end: range.end)
let result: FDB.KeyValuesResult = try await fdb.get(subspace: subspace)

// although call below is not equal to above one because `key(subspace:)` overload implicitly loads range
// this one will load bare subspace key
let result: FDB.KeyValuesResult = try await fdb.get(key: subspace)

result.records.forEach {
    dump("\($0.key) - \($0.value)")
}

Clearing values

Clearing (removing, deleting, you name it) records is simple as well.

try await fdb.clear(key: childSubspace["concrete_record"])

// OR
try await fdb.clear(key: rootSubspace["child"]["subspace"]["concrete_record"])

// OR EVEN
try await fdb.clear(key: rootSubspace["child", "subspace", "concrete_record"])

// OR EVEN (this is not OK, but still possible :)
try await fdb.clear(key: rootSubspace["child", FDB.Null, FDB.Tuple("foo", "bar"), "concrete_record"])

// clears whole subspace, including "concrete_record" key
try await fdb.clear(range: childSubspace.range)

Atomic operations

FoundationDB also supports atomic operations like ADD, AND, OR, XOR and stuff like that (please refer to docs). You can perform any of these operations with atomic(_ op:key:value:) method:

try await fdb.atomic(.add, key: key, value: 1)

Knowing that most popular atomic operation is increment (or decrement), I added handy syntax sugar:

try await fdb.increment(key: key)

// OR returning incremented value, which is always Int64
let result: Int64 = try await fdb.increment(key: key)

// OR
let result = try await fdb.increment(key: key, value: 2)

However, keep in mind that example above isn't atomic anymore.

And decrement, which is just a proxy for increment(key:value:), just inverting the value:

let result = try await fdb.decrement(key: key)

// OR
let result = try await fdb.decrement(key: key, value: 2)

Transactions

All previous examples are utilizing FDB object methods which are implicitly transactional. If you would like to perform more than one operation within one transaction (and experience all delights of ACID), you should first begin transaction using begin() method on FDB object context and then do your stuff (just don't forget to commit() it in the end, by default transactions roll back if not committed explicitly, or after timeout of 5 seconds):

let transaction = try fdb.begin()

transaction.set(key: "someKey", value: someBytes)

try await transaction.commit()

// OR
transaction.reset()

// OR
transaction.cancel()

Or you can just leave transaction object in place and it resets & destroys itself on deinit. Consider it auto-rollback. Please refer to official docs on reset and cancel behaviour: https://apple.github.io/foundationdb/api-c.html#c.fdb_transaction_reset

It's not really necessary to commit readonly transaction though :)

Additionally you may set transaction options using transaction.setOption(_:) method:

let transaction: AnyFDBTransaction = ...
try transaction.setOption(.transactionLoggingEnable(identifier: "debuggable_transaction"))
try transaction.setOption(.snapshotRywDisable)

See Transaction+Options.swift for a complete list of options.

Conflicts, retries and withTransaction

Since FoundationDB is quite a transactional database, sometimes commits might not succeed due to serialization failures (more commonly and mistakenly known as deadlocks). This can happen when two or more transactions create overlapping conflict ranges. Or, simply speaking, when they try to access or modify same keys (unless they are not in snapshot read mode) at the same time. This is expected (and, in a way, welcomed) behaviour, because this is how ACID works.

In these [not-so-rare] cases transaction is allowed to be replayed again. How do you know if your transaction can be replayed? It's failed with a special error caseFDB.Error.transactionRetry. If your transaction is failed with this particular error, it means that the transaction has already been rolled back to its initial state and is ready to be executed again.

You can implement this retry logic manually or you can just use FDB instance method withTransaction. Following example should be self-explanatory:

let maybeString: String? = try await fdb.withTransaction { transaction in
    guard let bytes: Bytes = try await transaction.get(key: key) else {
        return nil
    }
    try await transaction.commit()
    return String(bytes: bytes, encoding: .ascii)
}

Thus your block of code will be gently retried until transaction is successfully committed (or until databases decides that it's been retried enough times and it's time to let it go).

Writing to Versionstamped Keys

As a special type of atomic operation, values can be written to special keys that are guaranteed to be unique. These keys make use of an incomplete versionstamp within their tuples, which will be completed by the underlying cluster when data is written. The Versionstamp that was used within a transaction can then be retrieved so it can be referenced elsewhere.

An incomplete versionstamp can be created and added to tuples using the FDB.Versionstamp() initializer. The userData field is optional, and serves to further order keys if multiple are written within the same transaction.

Within a transaction's block, the set(versionstampedKey:value:) method can be used to write to keys with incomplete versionstamps. This method will search the key for an incomplete versionstamp, and if one is found, will flag it to be replaced by a complete versionstamp once it's written to the cluster. If an incomplete versionstamp was not found, a FDB.Error.missingIncompleteVersionstamp error will be thrown.

If you need the complete versionstamp that was used within the key, you can call getVersionstamp() before the transaction is committed. Note that this method must be called within the same transaction that a versionstamped key was written in, otherwise it won't know which versionstamp to return. Also note that this versionstamp does not include any user data that was associated with it, since it will be the same versionstamp no matter how many versionstamped keys were written.

let keyWithVersionstampPlaceholder = self.subspace[FDB.Versionstamp(userData: 42)]["anotherKey"]
let valueToWrite: String = "Hello, World!"

var versionstamp: FDB.Versionstamp = try await fdb.withTransaction { transaction in
    try transaction.set(versionstampedKey: keyWithVersionstampPlaceholder, value: Bytes(valueToWrite.utf8))
    try await transaction.commit()
    return try await transaction.getVersionstamp()
}

versionstamp.userData = 42
let actualKey = self.subspace[versionstamp]["anotherKey"]

// ... return it to user, save it as a reference to another entry, etc…

Complete example

let key = FDB.Subspace("1337")["322"]

let resultString: String = try await fdb.withTransaction { transaction in
    try transaction.setOption(.timeout(milliseconds: 5000))
    try transaction.setOption(.snapshotRywEnable)

    transaction.set(key: key, value: Bytes([1, 2, 3]))

    guard let bytes = try await transaction.get(key: key, snapshot: true) else {
        throw MyApplicationError.Something("Bytes are not bytes")
    }
    guard let string = String(bytes: bytes, encoding: .ascii) else {
        throw MyApplicationError.Something("String is not string")
    }

    try await transaction.commit()

    return string
}

print("My string is '\(resultString)'")

Debugging/logging

FDBSwift supports official community Swift-Log library with a custom backend LGNLog, therefore FDBSwift will use Logger.current logger (which may be bound to a TaskLocal logger via Logger.$current.withValue(contextLogger) { ... }) and obey log level and other LGNLogger config entries, see respective docs.

Troubleshooting

Package doesn't compile, something like Undefined symbols for architecture and tons of similar crap. Send help.

You haven't properly installed pkg-config for FoundationDB, see Installation section.

Package does compile in macOS, but in runtime I'm getting error The bundle “FDBTests” couldn’t be loaded because it is damaged or missing necessary resources. Try reinstalling the bundle. What do?

Execute this magic command in console: install_name_tool -id /usr/local/lib/libfdb_c.dylib /usr/local/lib/libfdb_c.dylib.

Shoutout to @dimitribouniol and his marvelous investigation.

I'm getting strange error on second operation: API version already set. Should I rethink my life?

(Rethinking hence analyzing things is always good) You tried to create more than one instance of FDB class, which is a) prohibited b) not needed at all since one instance is just enough for any application (if not, consider horizontal scaling, FDB absolutely shouldn't be a bottleneck of your application). Strictly speaking, it's not very ok, there should be a way of creating more than one of FDB connection in a runtime, and I will definitely try to make it possible. Still, I don't think that FDB connection pooling is a good idea, it already does everything for you.

Warnings

Though I aim for full interlanguage compatibility of Tuple layer, I don't guarantee it. During development I refered to Python implementation, but there might be slight differences (like unicode string and byte string packing, see design doc on strings and my comments on that). In general it's should be quite compatible already. Probably one day I'll spend some time on ensuring packing compatibility, but that's not high priority for me.

TODOs

  • Enterprise support, vendor WSDL, rewrite on Java Scala Kotlin Java 10
  • Drop enterprise support, rewrite on golang using react-native (pretty sure it will be a thing by that time)
  • Blockchain? ICO? VR? AR?
  • Rehab
  • ✅ Proper errors
  • ✅ Transactions rollback
  • ✅ Tuples
  • ✅ Tuples pack
  • ✅ Tuples unpack
  • ✅ Integer tuples
  • ✅ Ranges
  • ✅ Subspaces
  • ✅ Atomic operations
  • ✅ Tests
  • ✅ Properly test on Linux
  • ✅ 🎉 Asynchronous methods (Swift-NIO)
  • ✅ More verbose
  • ✅ Even more verbose
  • ✅ Transaction options
  • ✅ Network options
  • ✅ Docblocks and built-in documentation
  • ✅ Auto transaction retry if allowed and appropriate
  • ✅ 🎉 Even morer verbose (Swift-Log)
  • ✅ The rest of tuple pack/unpack (only floats, I think?) (also Bool and UUID)
  • ✅ Adopt async/await, yeah, boiiiiiiiiii
  • ✅ Drop Swift-NIO support (not because it's something bad, but because it's not really necessary here anymore; we still love it tho)
  • More sugar for atomic operations
  • The rest of C API (watches?)
  • Directories
  • Drop VR support