lmdb-safe

A safe modern & performant C++ wrapper of LMDB.
Requires C++17, or C++11 + Boost.

LMDB is an outrageously fast key/value store with semantics that make it highly interesting for many applications. Of specific note, besides speed, is the full support for transactions and good read/write concurrency. LMDB is also famed for its robustness.. when used correctly.

The design of LMDB is elegant and simple, which aids both the performance and stability. The downside of this elegant design is a nontrivial set of rules that need to be followed to not break things. In other words, LMDB delivers great things but only if you use it exactly right. This is by conscious design.

Among the things to keep in mind when using LMDB natively:

Never open a database file more than once anywhere in your process
Never open more than one transaction within a thread
- .. unless they are all read-only and have MDB_NOTLS set
When opening a named database, no other threads may do that at the same time
Cursors within RO transactions need freeing, but cursors within RW transactions must not be freed.
A new transaction may indicate the database has grown, and you need to restart the transaction then.

Breaking these rules may cause no immediate errors, but can lead to silent data corruption, missing updates, or random crashes. Again, this is not an actual bug in LMDB, it means that LMDB expects you to use it according to its exact rules. And who are we to disagree?

The lmdb-safe library aims to deliver the full LMDB performance while programmatically making sure the LMDB semantics are adhered to, with very limited overhead.

Most common LMDB functionality is wrapped within this library but the native MDB handles are all available should you want to use functionality we did not (yet) cater for.

In addition, on top of lmdb-safe, a type-safe "Object Relational Mapping" interface is also available. This auto-generates indexes and allows for the insertion, deletion and iteration of objects.

Status

Fresh. If using this tiny library, be aware things might change rapidly. To use, add lmdb-safe.cc and lmdb-safe.hh to your project. In addition, add lmdb-typed.hh to use the ORM.

Philosophy

This library tries to not restrict your use of LMDB, nor make it slower, except on operations that should be rare. The native LMDB handles (Environment, DBI, Transactions & Cursors) are all available for your direct use if need be.

When using lmdb-safe, errors "that should never happen" are turned into exceptions. An error that merely indicates that a key can not be found is passed on as a regular LMDB error code.

Example

The following example has no overhead compared to native LMDB, but already exhibits several ways in which lmdb-safe automates LMDB constraints:

  auto env = getMDBEnv("./database", 0, 0600);
  auto dbi = env->openDB("example", MDB_CREATE);
  auto txn = env->getRWTransaction();

The first line requests an LMDB environment for a database hosted in ./database. Within LMDB, it is not allowed to open a database file more than once, not even from other threads, not even when using a different LMDB handle. getMDBEnv keeps a registry of LMDB environments, keyed to the exact inode & flags. If another part of your process requests access to the same inode, it will get the same environment. MDBEnv is threadsafe.

On the second line, a database is opened within our environment. The semantics of opening or creating a database within LMDB are tricky. With some loss of generality, MDBEnv::openDB will create a transaction for you to open the database, and close it too. Most of the time this is what you want. It is also possible to open a database within a transaction manually.

The third line opens a read/write transaction using the Resource Acquisition Is Initialization (RAII) technique. If txn goes out of scope, the transaction is aborted automatically. To commit or abort, use commit() or abort(), after which going out of scope has no further effect.

  txn->put(dbi, "lmdb", "great");

  string_view data;
  if(!txn->get(dbi, "lmdb", data)) {
    cout<< "Within RW transaction, found that lmdb = " << data <<endl;
  }
  else
    cout<<"Found nothing" << endl;

  txn->commit();

LMDB is so fast because it does not copy data unless it really needs to. Memory bandwidth is a huge determinant of performance on modern CPUs. This wrapper agrees, and using modern C++ makes it possible to seemlessly use 'views' on data without copying them. Using these techniques, the call to txn.put() sets the "lmdb" string to "great", without making additional copies.

We employ the same technique to request the value of "lmdb", which is made available to us as a read-only view, straight onto the memory mapped data on disk.

In the final line, we commit the transaction, after which it also becomes available for other threads and processes.

A slightly expanded version of this code can be found in basic-example.cc.

Input and output of values

The basic data unit of LMDB is MDB_val which describes a slab of memory. Within LMDB, MDB_val is used for both input and output. For safety purposes, in this library we split this up into MDBInValue and MDBOutValue. Once split, we can add some very convenient semantics to these classes.

For example, to store double values for 64 bit IDs:

  auto txn = env->getRWTransaction();
  uint64_t id=12345678901;
  double score=3.14159;
  txn->put(dbi, id, score);
  txn->commit();

Behind the scenes, the id and score values are wrapped by MDBInVal which converts these values into byte strings. To retrieve thise values works similary:

  auto txn = env->getRWTransaction();
  uint64_t id=12345678901;
  MDBOutValue val;

  txn->get(dbi, id, val);

  cout << "Score: " << val.get<double>() << "\n";

Note that on retrieval, we have to specify the type of the value stored. This allows the conversion back from a byte string into the native type. MDBOutValue also tests if the length of the data matches the type.

Details

The automatic conversion to and from the MDBVals is implemented strictly for:

Integer and floating point types
std::string
std::string_view

However, if you explicitly ask for it, it is also possible to serialize structs:

struct Coordinate
{
    double x,y;
};

C c{12.0, 13.0};

txn->put(dbi, MDBInVal::fromStruct(c), 12.0);

MDBOutVal res;
txn->get(dbi, MDBInVal::fromStruct(c), res);

auto c1 = res.get_struct<Coordinate>();

Cursors, transactions

This example shows how to use cursors and how to mix lmdb-safe with direct calls to mdb.

  auto env = getMDBEnv("./database", 0, 0600);
  auto dbi = env->openDB("huge", MDB_CREATE);
  auto txn = env->getRWTransaction();

  unsigned int limit=20000000;

This is the usual opening sequence.

  auto cursor=txn->getCursor(dbi);
  MDBOutVal key, data;
  int count=0;
  cout<<"Counting records.. "; cout.flush();
  while(!cursor.get(key, data, count ? MDB_NEXT : MDB_FIRST)) {
    count++;
  }
  cout<<"Have "<<count<<"!"<<endl;

This describes how we generate a cursor for the huge database and iterate over it to count the number of keys in there. We pass two LMDB native MDB_val structs to the cursor get function. These do not get copies of all the millions of potential keys in the huge database - they only contain pointers to that data. Because of this, we can count 20 million records in under a second (!).

  cout<<"Clearing records.. "; cout.flush();
  mdb_drop(*txn, dbi, 0); // clear records
  cout<<"Done!"<<endl;

Here we drop al keys from the database, which too happens nearly instantaneously. Note that we pass our txn (which is a class) to the native mdb_drop function which we did not wrap. This is possible because txn converts to an MDB_env* if needed.

  cout << "Adding "<<limit<<" values  .. "; cout.flush();
  for(unsigned int n = 0 ; n < limit; ++n) {
    txn->put(dbi, n, n);
  }
  cout <<"Done!"<<endl;
  cout <<"Calling commit.. "; cout.flush();
  txn->commit();
  cout<<"Done!"<<endl;

Here we add 20 million value & then commit the mdb_drop and the 20 million puts. All this happened in less than 20 seconds.

Had we created our database with the MDB_INTEGERKEY option and added the MDB_APPEND flag to txn.put, the whole process would have taken around 5 seconds.

lmdb-typed

The lmdb-safe interface may be safe in one sense, but it is still a key-value store, allowing the user to store any key and any value. Frequently we have specific needs: to store objects and find them using different keys. Doing so manually is cumbersome and error-prone, as all indexes (for rapid retrieval) need to be carefully maintained by hand.

Inspired by Boost MultiIndex, lmdb-typed builds on lmdb-safe to create, populate and use indexes for rapidly retrieving objects. As an example, let's say we want to store the following struct:

struct DNSResourceRecord
{
  string qname;          // index
  uint16_t qtype{0};
  uint32_t domain_id{0}; // index
  string content;
  uint32_t ttl{0};
  string ordername;      // index
  bool auth{true};
};

And we want to do so based on the qname, domain_id or ordername fields. First, we have to make sure DNSResourceRecord can serialize itself to a string:

template<class Archive>
void serialize(Archive & ar, DNSResourceRecord& g, const unsigned int version)
{
  ar & g.qtype;
  ar & g.qname;
  ar & g.content;
  ar & g.ttl;
  ar & g.domain_id;
  ar & g.ordername;
  ar & g.auth;
}

Next up, we need to define our "Object Relational Mapper":

  TypedDBI<DNSResourceRecord, 
           index_on<DNSResourceRecord, string,   &DNSResourceRecord::qname>,
           index_on<DNSResourceRecord, uint32_t, &DNSResourceRecord::domain_id>,
           index_on<DNSResourceRecord, string,   &DNSResourceRecord::ordername>
           > tdbi(getMDBEnv("./typed.lmdb", MDB_NOSUBDIR, 0600), "records");

This defines that we create a database called records in the file ./typed.lmdb. We also state that this database stores DNSResourceRecord objects, and that we want three indexes. Note that this syntax is reasonable similar to that used by Boost::MultiIndex.

Next up, we can insert some objects:

auto txn = tdbi.getRWTransaction();
DNSResourceRecord rr{"www.powerdns.com", 1, domain_id, "1.2.3.4", 0, "www"};
// populate rr
auto id = txn.insert(rr);
txn.commit();

Internally, the opening of tdbi above created four databases: records, records_0, records_1 and records_2. On insert, a serialized form of rr was stored in the records table, with the key containing the (assigned) id value.

In addition, in records_1, the qname was added as key, with the id field as value. And similarly for domain_id and ordername. So the indexes all point to the id field, which we can find in the records database.

To retrieve, we can use any of the indexes:

auto txn = tdbi.getROTransaction(); 
DNSResourceRecord rr;
txn.get(id, rr);
txn.get<0>("www.powerdns.com", rr);
txn.get<1>(domain_id, rr);
txn.get<2>("www", rr);

As long as we inserted only the one DNSResourceRecord from above, all four get calls find the same rr.

In the more interesting case where we inserted more DNS records, we could iterate over all items with domain_id = 4 as follows:

for(auto iter = txn.find<1>(4): iter != txn.end(); ++iter) {
    cout << iter->qname << "\n";
}

To delete an item, use txn.del(12), which will remove the record with id 12 from the main database and also from all the indexes.

ahupowerdns / lmdb-safe

readme