Raw SST File Iterator & reader

[swamirishi]

Raw SST File Reader for reading tombstone entries from sst file along with sequence number & type of the data in rocksdb

[jowlyzhang]

@swamirishi Thank you for this work! Just curious, is the motivation for this class to be able to read the sequence number and type of each entry from a table file? In that case, RocksDB's SstFileDumper and SstFileReader can be augmented to surface these two items too. For example, in SstFileDumper:

rocksdb/table/sst_file_dumper.cc

Line 500 in cb4f438

Status pik_status = ParseInternalKey(key, &ikey, true /* log_err_key */);

ikey is a parsed internal key, you can get the sequence number and type with ikey.sequence, ikey.type. And maybe add some command line option for the sst_dump tool to conditionally print these two things out.

[swamirishi]

@swamirishi Thank you for this work! Just curious, is the motivation for this class to be able to read the sequence number and type of each entry from a table file? In that case, RocksDB's SstFileDumper and SstFileReader can be augmented to surface these two items too. For example, in SstFileDumper:

rocksdb/table/sst_file_dumper.cc

Line 500 in cb4f438

Status pik_status = ParseInternalKey(key, &ikey, true /* log_err_key */);

ikey is a parsed internal key, you can get the sequence number and type with ikey.sequence, ikey.type. And maybe add some command line option for the sst_dump tool to conditionally print these two things out.

@jowlyzhang In Ozone we use rocksdb for metadata store. We implemented snapshots in ozone relying on the rocksdb checkpoint functionality. In order to perform efficient snapshot diffs we currently need all the tombstone entries that are written to the sst files to figure out the keys that have changed over the course of multiple checkpoints. Currently we are patching up the rocksdb code and are building this particular tool and wrote our own jni layer to access the tombstone entries and the sequence number. It would be a really great functionality on the sst file reader. Does this seem like a valid usecase and ask from a feature perspective? I can work on it to augment the sst file reader to be able to do this by adding a flag on read options. You can take a look at this PR to get a better understanding apache/ozone#6182

[swamirishi]

@swamirishi Thank you for this work! Just curious, is the motivation for this class to be able to read the sequence number and type of each entry from a table file? In that case, RocksDB's SstFileDumper and SstFileReader can be augmented to surface these two items too. For example, in SstFileDumper:

rocksdb/table/sst_file_dumper.cc

Line 500 in cb4f438

Status pik_status = ParseInternalKey(key, &ikey, true /* log_err_key */);

ikey is a parsed internal key, you can get the sequence number and type with ikey.sequence, ikey.type. And maybe add some command line option for the sst_dump tool to conditionally print these two things out.

Currently the db_iter skips non user keys

rocksdb/db/db_iter.cc

Line 289 in 003197f

bool DBIter::FindNextUserEntryInternal(bool skipping_saved_key,

, currently the sst file reader is tightly coupled with this.

[jowlyzhang]

Currently the db_iter skips non user keys

rocksdb/db/db_iter.cc

Line 289 in 003197f

bool DBIter::FindNextUserEntryInternal(bool skipping_saved_key,

, currently the sst file reader is tightly coupled with this.

sst file reader should be returning a table iterator that iterates the table, not a DBIter:

rocksdb/table/sst_file_reader.cc

Line 90 in cb4f438

auto internal_iter = r->table_reader->NewIterator(

For block based table, this would be a BlockBasedTableIterator:

rocksdb/table/block_based/block_based_table_iterator.h

Line 24 in cb4f438

BlockBasedTableIterator(

This iterator iterates the whole table file, tombstones are surfaced too.

[swamirishi]

Currently the db_iter skips non user keys

rocksdb/db/db_iter.cc

Line 289 in 003197f

bool DBIter::FindNextUserEntryInternal(bool skipping_saved_key,

, currently the sst file reader is tightly coupled with this.

sst file reader should be returning a table iterator that iterates the table, not a DBIter:

rocksdb/table/sst_file_reader.cc

Line 90 in cb4f438

auto internal_iter = r->table_reader->NewIterator(

For block based table, this would be a BlockBasedTableIterator:

rocksdb/table/block_based/block_based_table_iterator.h

Line 24 in cb4f438

BlockBasedTableIterator(

This iterator iterates the whole table file, tombstones are surfaced too.

rocksdb/table/sst_file_reader.cc

Line 94 in cb4f438

res->SetIterUnderDBIter(internal_iter);

makes it a db_iter

[jowlyzhang]

rocksdb/table/sst_file_reader.cc

Line 94 in cb4f438

res->SetIterUnderDBIter(internal_iter);

makes it a db_iter

I see, so you need an iterator to pragmatically iterate the raw sst file to get the tombstone. So you want to define a public iterator class. Would separate tool like sst_dump work for your flow? This tool can be augmented to print out the type and sequence number, you would need to parse its output.

We have a feature to allow users to collect table properties, it sounds like it can work for your use case. You can use it to pragmatically collect the tombstones from each SST file. You can define a TablePropertiesCollectorFactory

rocksdb/include/rocksdb/table_properties.h

Line 153 in cb4f438

class TablePropertiesCollectorFactory : public Customizable {

This factory is responsible for creating a TablePropertiesCollector

rocksdb/include/rocksdb/table_properties.h

Line 94 in cb4f438

virtual ~TablePropertiesCollector() {}

When each SST file is being created, this TablePropertiesCollector::AddUserKey interface will be invoked with the corresponding user key, entry type, sequence number, etc.

rocksdb/include/rocksdb/table_properties.h

Lines 110 to 112 in cb4f438

	virtual Status AddUserKey(const Slice& key, const Slice& value,
	EntryType /*type*/, SequenceNumber /*seq*/,
	uint64_t /*file_size*/) {

You can define your own TablePropertiesCollector to collect these tombstone entries. What RocksDB will do is to call the TablePropertiesCollector::GetReadableProperties interface to get a user defined table properties and persist it to the corresponding SST file. You may want to piggyback this to save this whole info directly in the SST file, or you can just save this info somewhere else. And persist an index to look for this info in the corresponding SST file.

Later on when this SST file needs to be processed, you can use SstFileReader::GetTableProperties API to get this user defined property and collect those tomstone info.

[swamirishi]

you want to define a public iterator class. Would separate tool like sst_dump work for your flow? This tool can be augmented to print out the type and sequence number, you would need to parse its output.

We actually wanted a jni which would let us iterate through the sst file and get the key,value, sequenceNumber for each of the records including the tombstone entries from the sst file. We are not particularly looking to get the table properties.

[jowlyzhang]

We actually wanted a jni which would let us iterate through the sst file and get the key,value, sequenceNumber for each of the records including the tombstone entries from the sst file. We are not particularly looking to get the table properties.

This function TablePropertiesCollector::AddUserKey will be invoked for each record with its key, value, sequence number, including tombstone entries for a SST file when it's created. So you are kind of iterating through the whole sst file. The only issue is, within this function, you do not know which SST file this is for. If you just need "all the tombstone entries that are written to the sst files to figure out the keys that have changed over the course of multiple checkpoints", you can collect this info and persist it in a file, say "0011.tombstone_entries.txt". And you can persist a user defined table properties: "tomstone_file_location:0011.tombstone_entries.txt" in the sst file. Later on when you need to apply that logic related to checkpoints, you can get this user defined table properties, and read the corresponding txt file for all the information you need. How you create this file, parse this file, iterate this file will be completely outside of RocksDB. And wouldn't need any RocksDB interface change.

[jowlyzhang]

We actually wanted a jni which would let us iterate through the sst file and get the key,value, sequenceNumber for each of the records including the tombstone entries from the sst file. We are not particularly looking to get the table properties.

RocksDB have some implementations that can be leveraged to create such a raw table iterator. Some of the changes in this PR are not required, I have prototyped a simpler change based on these implementations to achieve similar effect. It's at: main...jowlyzhang:rocksdb:internal_key_interator

Let me know if this looks OK to you, I can work on checking it in.

[swamirishi]

We actually wanted a jni which would let us iterate through the sst file and get the key,value, sequenceNumber for each of the records including the tombstone entries from the sst file. We are not particularly looking to get the table properties.

RocksDB have some implementations that can be leveraged to create such a raw table iterator. Some of the changes in this PR are not required, I have prototyped a simpler change based on these implementations to achieve similar effect. It's at: main...jowlyzhang:rocksdb:internal_key_interator

Let me know if this looks OK to you, I can work on checking it in.

Yup this change definitely solves my purpose

[jowlyzhang]

Hello @swamirishi, I have put together a PR to support raw table iterator in #12385

Please feel free to comment if any functionality is missing. This PR doesn't include the jni wrapper yet. I will add those after we have settled that the functionality added there are sufficient.

[jowlyzhang]

Hello @swamirishi, since we have merged #12385, shall we close this PR?

[swamirishi]

Hello @swamirishi, since we have merged #12385, shall we close this PR?

Yeah this can be closed in favour of #12385