LogDevice API
Public Member Functions | Static Public Member Functions | Public Attributes | List of all members
facebook::logdevice::Reader Class Referenceabstract

Public Member Functions

virtual int startReading (logid_t log_id, lsn_t from, lsn_t until=LSN_MAX, const ReadStreamAttributes *attrs=nullptr)=0
 
virtual int stopReading (logid_t log_id)=0
 
virtual bool isReading (logid_t log_id) const =0
 
virtual bool isReadingAny () const =0
 
virtual int setTimeout (std::chrono::milliseconds timeout)=0
 
virtual ssize_t read (size_t nrecords, std::vector< std::unique_ptr< DataRecord >> *data_out, GapRecord *gap_out)=0
 
virtual void waitOnlyWhenNoData ()=0
 
virtual void withoutPayload ()=0
 
virtual void forceNoSingleCopyDelivery ()=0
 
virtual void includeByteOffset ()=0
 
virtual void doNotSkipPartiallyTrimmedSections ()=0
 
virtual int isConnectionHealthy (logid_t) const =0
 
virtual void doNotDecodeBufferedWrites ()=0
 
virtual ~Reader ()
 

Static Public Member Functions

static lsn_t nextFromLsnWhenStuck (lsn_t stuck_lsn=LSN_INVALID, lsn_t tail_lsn=LSN_INVALID)
 

Public Attributes

const std::chrono::milliseconds MAX_TIMEOUT {(1LL << 31) - 1}
 

Constructor & Destructor Documentation

◆ ~Reader()

virtual facebook::logdevice::Reader::~Reader ( )
inlinevirtual

Note that, unless all reading is stopped prior to destruction, the destructor may block for some time while all reading gets stopped. This will typically finish quickly in applications but it does involve interthread communication for each log.

Member Function Documentation

◆ doNotDecodeBufferedWrites()

virtual void facebook::logdevice::Reader::doNotDecodeBufferedWrites ( )
pure virtual

Instructs the Reader instance to pass through blobs created by BufferedWriter.

By default (if this method is not called), Reader automatically decodes blobs written by BufferedWriter and yields original records as passed to BufferedWriter::append(). If this method is called, BufferedWriteDecoder can be used to decode the blobs.

◆ doNotSkipPartiallyTrimmedSections()

virtual void facebook::logdevice::Reader::doNotSkipPartiallyTrimmedSections ( )
pure virtual

If called, when reading a section of the log that has been partially trimmed, the reader will deliver whatever records are still available, which (because of LogDevice's distributed and nondeterministic nature) results in an interleaved stream of records and TRIM gaps, which is undesirable in some cases.

The default behaviour is to deliver a large trim gap for the entire section.

See doc/partially-trimmed.md for a detailed explanation.

◆ forceNoSingleCopyDelivery()

virtual void facebook::logdevice::Reader::forceNoSingleCopyDelivery ( )
pure virtual

If called, disable the single copy delivery optimization even if the log is configured to support it. Each data record will be sent by all storage nodes that have a copy instead of exactly one. This greatly increases read availability at the cost of higher network bandwith and cpu usage.

Only affects subsequent startReading() calls.

◆ includeByteOffset()

virtual void facebook::logdevice::Reader::includeByteOffset ( )
pure virtual

If called, data records read by this Reader will start including approximate amount of data written to given log up to current record once it become available to Reader.

The value itself stored in DataRecord::attrs::byte_offset. Set as BYTE_OFFSET_INVALID if unavailable to Reader yet.

Only affects subsequent startReading() calls.

◆ isConnectionHealthy()

virtual int facebook::logdevice::Reader::isConnectionHealthy ( logid_t  ) const
pure virtual

Checks if the connection to the LogDevice cluster for a log appears healthy. When a read() call times out, this can be used to make an informed guess whether this is because there is no data or because there a service interruption.

NOTE: this is not 100% accurate but will expose common issues like losing network connectivity.

Returns
On success, returns 1 if the connection appears healthy or 0 if there are issues talking to the cluster. On error returns -1 and sets err to NOTFOUND (not reading given log).

◆ isReading()

virtual bool facebook::logdevice::Reader::isReading ( logid_t  log_id) const
pure virtual

Checks if a log is being read. Can be used to find out if the end of the log was reached (for a log that was being read).

◆ isReadingAny()

virtual bool facebook::logdevice::Reader::isReadingAny ( ) const
pure virtual

Checks if any log is being read. Can be used to find out if the end was reached for all logs that were being read.

◆ nextFromLsnWhenStuck()

static lsn_t facebook::logdevice::Reader::nextFromLsnWhenStuck ( lsn_t  stuck_lsn = LSN_INVALID,
lsn_t  tail_lsn = LSN_INVALID 
)
static

Get next lowest recommended LSN to read from when servers appear stuck.

To force progress when stuck at stuck_lsn, call stopReading(), then call startReading() from nextFromLsnWhenStuck(stuck_lsn).

NOTE: Forcing progress is likely to miss some records between stuck_lsn and nextFromLsnWhenStuck(stuck_lsn). The records will become available eventually (once server-side issues have been resolved), but the only way to read them is to re-read the section from stuck_lsn to nextFromLsnWhenStuck(stuck_lsn) later.

Parameters
stuck_lsnLSN at which the reader got stuck; that is, the last LSN the reader tried to read or managed to read (+-1 does not affect the output of this function).
tail_lsnTail LSN, if known, as returned by Client::getTailLSN() or Client::findTime().
Returns
Recommended LSN to use as from_lsn in startRead() to make progress, no less than stuck_lsn.

◆ read()

virtual ssize_t facebook::logdevice::Reader::read ( size_t  nrecords,
std::vector< std::unique_ptr< DataRecord >> *  data_out,
GapRecord *  gap_out 
)
pure virtual

Attempts to read a batch of records.

The call either delivers 0 or more (up to nrecords) data records, or one gap record.

The call returns when any of this is true:

  • nrecords records have been delivered
  • there are no more records to deliver at the moment and the timeout specified by setTimeout() has been reached
  • there are no more records to deliver at the moment and waitOnlyWhenNoData() was called
  • a gap in sequence numbers is encountered
  • until LSN for some log was reached
  • not reading any logs, possibly because the ends of all logs have been reached (returns 0 quickly)

Note that even in the case of an infinite timeout, the call may deliver less than nrecords data records when a gap is encountered. The next call to read() will deliver the gap.

Waiting will not be interrupted if a signal is delivered to the thread.

Example usage: std::vector<std::unique_ptr<DataRecord> > records; GapRecord gap; ssize_t nread = reader->read(100, &records, &gap); if (nread >= 0) { for (int i=0; i<nread; ++i) { // process *records[i] } } else { assert(err == E::GAP); // process gap }

Parameters
nrecordslimit on number of records to return data_out pointer to vector to append data records to gap_out pointer to a single GapRecord instance, populated when there is a gap in sequence numbers
Returns
Returns the number of records delivered (between 0 and nrecords), or -1 if there was a gap. If >= 0, that many data records were appended to the vector. If -1 then logdevice::err is set to E::GAP, and *gap_out is filled with information about the gap.

◆ setTimeout()

virtual int facebook::logdevice::Reader::setTimeout ( std::chrono::milliseconds  timeout)
pure virtual

Sets the limit on how long read() calls may wait for records to become available. A timeout of -1 means no limit (infinite timeout). A timeout of 0 means no waiting (nonblocking reads).

Default is no limit.

The maximum timeout is 2^31-1 milliseconds (about 24 days). If a timeout larger than that is passed in, it will be capped.

Returns
0 on success, -1 if the parameter was invalid

◆ startReading()

virtual int facebook::logdevice::Reader::startReading ( logid_t  log_id,
lsn_t  from,
lsn_t  until = LSN_MAX,
const ReadStreamAttributes *  attrs = nullptr 
)
pure virtual

Start reading a log. Similar to AsyncReader::startReading().

Any one log can only be read once by a single Reader. If this method is called for the same log multiple times, it restarts reading, optionally at a different point.

Returns
On success, returns 0. On failure, -1 is returned and logdevice::err is set to: NOBUFS if request could not be enqueued because a buffer space limit was reached INVALID_PARAM if from > until or the record callback was not specified. NOTFOUND if the log doesn't exist SHUTDOWN the logdevice::Client instance was destroyedA. INTERNAL an internal error has been detected, check logs. TOOMANY exceeded limit on number of logs that had been specified the Reader was created.

◆ stopReading()

virtual int facebook::logdevice::Reader::stopReading ( logid_t  log_id)
pure virtual

Stop reading a log.

Returns
On success, returns 0. On failure, -1 is returned and logdevice::err is set to: [any of the error codes from AsyncReader::stopReading()] NOTFOUND log is not being read

◆ waitOnlyWhenNoData()

virtual void facebook::logdevice::Reader::waitOnlyWhenNoData ( )
pure virtual

If called, whenever read() can return some records but not the number requested by the caller, it will return the records instead of waiting for more.

Example:

  • Caller calls read(100, ...) asking for 100 data records.
  • Only 20 records are immediately available.
  • By default, read() would wait until 80 more records arrive or the timeout expires. This makes sense for callers that can benefit from reading and processing batches of data records.
  • If this method was called before read(), it would return the 20 records without waiting for more. This may make sense for cases where latency is more important.

◆ withoutPayload()

virtual void facebook::logdevice::Reader::withoutPayload ( )
pure virtual

If called, data records read by this Reader will not include payloads.

This makes reading more efficient when payloads are not needed (they won't be transmitted over the network).

Only affects subsequent startReading() calls.


The documentation for this class was generated from the following file: