I am using questdb (embedded) to store a bunch of time series.
I would like to run my storage method inside a parallel stream, but I don't know if TableWriter is thread-safe.
Here is the code:
SqlExecutionContextImpl ctx = new SqlExecutionContextImpl(engine, 1);
try (TableWriter writer = engine.getWriter(ctx.getCairoSecurityContext(), name, "writing")) {
tickerData.stream().parallel().forEach(
r -> {
Instant i = r.getDateTime("DateTime")
.atZone(EST)
.toInstant();
long ts = TimestampFormatUtils.parseTimestamp(i.toString());
TableWriter.Row row = writer.newRow(ts);
row.putDouble(0, r.getDouble("x1"));
row.putDouble(1, r.getDouble("x2"));
row.putDouble(2, r.getDouble("y1"));
row.putDouble(3, r.getDouble("y2"));
row.putDouble(4, r.getDouble("z"));
row.append();
writer.commit();
} catch (NumericException ex) {
log.error("Cannot parse the date {}", r.getDateTime("DateTime"));
} catch (Exception ex) {
log.error("Cannot write to table {}!", name, ex);
}
});
}
This throws all sort of errors, is there a way to make the storage process parallel?
Thanks,
Juan
The short answer is TableWriter is not thread safe. You will be responsible to not use it in parallel threads.
A bit longer answer is that even in stand alone QuestDB parallel writing is restricted. It is only possible from multiple ILP connections at the moment.
Im basically facing a blocking problem.
I have my server coded based on C++ Boost.ASIO using 8 threads since the server has 8 logical cores.
My problem is a thread may face 0.2~1.5 seconds of blocking on a MySQL query and I honestly don't know how to go around that since MySQL C++ Connector does not support asynchronous queries, and I don't know how to design the server "correctly" to use multiple threads for doing the queries.
This is where I'm asking for opinions of what to do in this case.
Create 100 threads for async' query sql?
Could I have an opinion from experts about this?
Okay, the proper solution to this would be to extend Asio and write a mysql_service implementation to integrate this. I was almost going to find out how this is done right away, but I wanted to get started using an "emulation".
The idea is to have
your business processes using an io_service (as you are already doing)
a database "facade" interface that dispatches async queries into a different queue (io_service) and posts the completion handler back onto the business_process io_service
A subtle tweak needed here you need to keep the io_service on the business process side from shutting down as soon as it's job queue is empty, since it might still be awaiting a response from the database layer.
So, modeling this into a quick demo:
namespace database
{
// data types
struct sql_statement { std::string dml; };
struct sql_response { std::string echo_dml; }; // TODO cover response codes, resultset data etc.
I hope you will forgive my gross simplifications :/
struct service
{
service(unsigned max_concurrent_requests = 10)
: work(io_service::work(service_)),
latency(mt19937(), uniform_int<int>(200, 1500)) // random 0.2 ~ 1.5s
{
for (unsigned i = 0; i < max_concurrent_requests; ++i)
svc_threads.create_thread(boost::bind(&io_service::run, &service_));
}
friend struct connection;
private:
void async_query(io_service& external, sql_statement query, boost::function<void(sql_response response)> completion_handler)
{
service_.post(bind(&service::do_async_query, this, ref(external), std::move(query), completion_handler));
}
void do_async_query(io_service& external, sql_statement q, boost::function<void(sql_response response)> completion_handler)
{
this_thread::sleep_for(chrono::milliseconds(latency())); // simulate the latency of a db-roundtrip
external.post(bind(completion_handler, sql_response { q.dml }));
}
io_service service_;
thread_group svc_threads; // note the order of declaration
optional<io_service::work> work;
// for random delay
random::variate_generator<mt19937, uniform_int<int> > latency;
};
The service is what coordinates a maximum number of concurrent requests (on the "database io_service" side) and ping/pongs the completion back onto another io_service (the async_query/do_async_query combo). This stub implementation emulates latencies of 0.2~1.5s in the obvious way :)
Now comes the client "facade"
struct connection
{
connection(int connection_id, io_service& external, service& svc)
: connection_id(connection_id),
external_(external),
db_service_(svc)
{ }
void async_query(sql_statement query, boost::function<void(sql_response response)> completion_handler)
{
db_service_.async_query(external_, std::move(query), completion_handler);
}
private:
int connection_id;
io_service& external_;
service& db_service_;
};
connection is really only a convenience so we don't have to explicitly deal with various queues on the calling site.
Now, let's implement a demo business process in good old Asio style:
namespace domain
{
struct business_process : id_generator
{
business_process(io_service& app_service, database::service& db_service_)
: id(generate_id()), phase(0),
in_progress(io_service::work(app_service)),
db(id, app_service, db_service_)
{
app_service.post([=] { start_select(); });
}
private:
int id, phase;
optional<io_service::work> in_progress;
database::connection db;
void start_select() {
db.async_query({ "select * from tasks where completed = false" }, [=] (database::sql_response r) { handle_db_response(r); });
}
void handle_db_response(database::sql_response r) {
if (phase++ < 4)
{
if ((id + phase) % 3 == 0) // vary the behaviour slightly
{
db.async_query({ "insert into tasks (text, completed) values ('hello', false)" }, [=] (database::sql_response r) { handle_db_response(r); });
} else
{
db.async_query({ "update * tasks set text = 'update' where id = 123" }, [=] (database::sql_response r) { handle_db_response(r); });
}
} else
{
in_progress.reset();
lock_guard<mutex> lk(console_mx);
std::cout << "business_process " << id << " has completed its work\n";
}
}
};
}
This business process starts by posting itself on the app service. It then does a number of db queries in succession, and eventually exits (by doing in_progress.reset() the app service is made aware of this).
A demonstration main, starting 10 business processes on a single thread:
int main()
{
io_service app;
database::service db;
ptr_vector<domain::business_process> bps;
for (int i = 0; i < 10; ++i)
{
bps.push_back(new domain::business_process(app, db));
}
app.run();
}
In my sample, business_processes don't do any CPU intensive work, so there's no use in scheduling them across CPU's, but if you wanted you could easily achieve this, by replacing the app.run() line with:
thread_group g;
for (unsigned i = 0; i < thread::hardware_concurrency(); ++i)
g.create_thread(boost::bind(&io_service::run, &app));
g.join_all();
See the demo running Live On Coliru
I'm not a MySQL guru, but the following is generic multithreading advice.
Having NumberOfThreads == NumberOfCores is appropriate when none of the threads ever block and you are just splitting the load over all CPUs.
A common pattern is to have multiple threads per CPU, so one is executing while another is waiting on something.
In your case, I'd be inclined to set NumberOfThreads = n * NumberOfCores where 'n' is read from a config file, a registry entry or some other user-settable value. You can test the system with different values of 'n' to fund the optimum. I'd suggest somewhere around 3 for a first guess.
I am currently running into several problems while threading a couple different processes with boost. I am fairly new to threading so bear with me. Hopefully the formatting is good enough as well.
I have two threads defined.
boost::thread m_thread1;
boost::thread m_thread2;
Associated with these threads are three futures and two packaged_tasks
boost::packaged_task<bool> ptaskBool;
boost::packaged_task<double> ptaskDouble;
boost::unique_future<bool> futureBool;
boost::unique_future<double> futureDouble1;
boost::unique_future<double> futureDouble2;
Both of these threads are defined inside a while loop
while(running)
{
//some code
if(!thread1running)
{
ptaskBool = boost::packaged_task<bool>(boost::bind(/*some objects function*/));
futureBool = ptaskBool.get_future();
m_thread1 = boost::thread(boost::move(ptaskBool));
thread1running = true;
}
if(!thread2running && initial)
{
ptaskDouble = boost::packaged_task<bool>(boost::bind(/*some objects function*/));
futureDouble = ptaskDouble.get_future();
m_thread2 = boost::thread(boost::move(ptaskDouble));
thread2running = true;
}
else if(!thread2running)
{
ptaskDouble = boost::packaged_task<bool>(boost::bind(/*some different function*/));
futureDouble2 = ptaskDouble.get_future();
m_thread2 = boost::thread(boost::move(ptaskDouble));
initial = true;
thread2running = true;
}
This is a simplification of what I have but it is very similar. The next code I have is for the futures.
if(futureBool.is_ready())
{
m_bool = futureBool.get();
thread1running = false;
}
if(futureDouble.is_ready())
{
m_double = futureDouble.get();
std::cout << "Result: " << m_double << std::endl;
thread2running = false;
//have tried several things like a join here to no avail
}
if(futureDouble2.is_ready())
{
m_double = futureDouble2.get();
std::cout << "Result: " << m_double << std::endl;
thread2running = false;
}
}//end while
What I meant by sessions in the title is that there is some session handling in the background to do with telling the threads when and when not to be created.
There is other code that does some processing and flag checking but it is not relevant to my question. I am clearly misunderstanding something that is going on. The first time each thread is run, they complete their tasks. The second time the program just crashes.
I have had luck with making each thread variable a pointer and defining a new thread every time instead of redefining each thread. This worked but I was unable to delete each thread after completion. What am I not understanding here? What is the thread actually doing after the function returns a result?
auto_ptr is a good way to ensure your threads will be deleted at the end. Also is unique_ptr from C++11.
According to: http://www.cplusplus.com
Note: This class template (auto_ptr) is deprecated as of C++11. unique_ptr is a new facility with a similar functionality, but with improved security (no fake copy assignments), added features (deleters) and support for arrays. See unique_ptr for additional information.
See Using auto_ptr Effectively for examples using auto_ptr.
So the mongo c++ documentation says
On a failover situation, expect at least one operation to return an
error (throw an exception) before the failover is complete. Operations
are not retried
Kind of annoying, but that leaves it up to me to handle a failed operation. Ideally I would just like the application to sleep for a few seconds (app is single threaded). And retry with the hopes that a new primary mongod is established. In the case of a second failure, well I take it the connection is truly messed up and I just want to thrown an exception.
Within my MongodbManager class this means all operations have this kind of double try/catch block set up. I was wondering if there is a more elegant solution?
Example method:
template <typename T>
std::string
MongoManager::insert(std::string ns, T object)
{
mongo::BSONObj = convertToBson(object);
std::string result;
try {
connection_->insert(ns, oo); //connection_ = shared_ptr<DBClientReplicaSet>
result = connection_->getLastError();
lastOpSucceeded_ = true;
}
catch (mongo::SocketException& ex)
{
lastOpSucceeded_ = false;
boost::this_thread::sleep( boost::posix_time::seconds(5) );
}
// try again?
if (!lastOpSucceeded_) {
try {
connection_->insert(ns, oo);
result = connection_->getLastError();
lastOpSucceeded_ = true;
}
catch (mongo::SocketException& ex)
{
//do some clean up, throw exception
}
}
return result;
}
That's indeed sort of how you need to handle it. Perhaps instead of having two try/catch blocks I would use the following strategy:
keep a count of how many times you have tried
create a while loop with as terminator (count < 5 && lastOpSucceeded)
and then sleep with pow(2,count) to sleep more in every iteration.
And then when all else fails, bail out.
I have a method (from a third party lib)
bool READ_DB(Connection* con, long value);
void TAMLTradeProcessor::CreateThreads() {
long nThreads = 15; // we can configure this
// set up and run the threads
HANDLE* pWaitHandles = new HANDLE[nThreads];
CThreadInfoData* pTid = new CThreadInfoData[nThreads];
UINT nRunningThreads = 0;
long lSharedIndex = -1;
// Initialise data blocks
int i;
for (i = 0; i < nThreads; i++)
{
pTid[i].m_bRunning = true;
pTid[i].m_pnCurrentIndexPosition = &lSharedIndex; // common index
pTid[i].m_pDbConn = new CDatabaseConnection();
pTid[i].m_hThread = (HANDLE )_beginthreadex(NULL,0,ThreadCB,&pTid[i],0,&pTid[i].m_nThreadId);
...
}
It reads data off the database using the connection I pass in and matching the query for that specific value.
I have a huge list of values so I created multiple threads that retrieves values off the list and call the method, in other word I am retrieving the data in parallel using mutiple DB connection.
ThreadCB will call READ_DB.
At the moment I have created the threads myself and I have created 15 of them...just a casual number.
Is there a better way of doing this using Windows ThreadPool API?
In other words if I need to run the same DB query over and over gain for different values (but I can only use one value at time) what is the best approach?