Support state history log splitting #9277
Conversation
|
As described there seems to be no way to share state history files between different nodeos instances. Obviously each nodeos needs to keep very recent stuff on its own, but once you want something older than say, a few days, then it would be nice to store the files on a shared drive (NFS, s3fuse, etc) and have mutliple nodeos processes pull from them. This way you get scalability on nodeos by having multiple, but without using multiple times the storage. |
For gcc 8 compatibility
| // Skip if not a file | ||
| if (!bfs::is_regular_file(p->status())) | ||
| continue; | ||
| // skip if it's not match blocks-*-*.log |
There was a problem hiding this comment.
is this comment true for this generic method?
| auto existing_itr = collection.find(log.first_block_num()); | ||
| if (existing_itr != collection.end()) { | ||
| if (log.last_block_num() <= existing_itr->second.last_block_num) { | ||
| wlog("${log_path} contains the overlapping range with ${existing_path}.log, droping ${log_path} " |
| if (existing_itr != collection.end()) { | ||
| if (log.last_block_num() <= existing_itr->second.last_block_num) { | ||
| wlog("${log_path} contains the overlapping range with ${existing_path}.log, droping ${log_path} " | ||
| "from catelog", |
| } | ||
|
|
||
| bool index_matches_data(const bfs::path& index_path, const LogData& log) const { | ||
| if (bfs::exists(index_path) && bfs::file_size(index_path) / sizeof(uint64_t) != log.num_blocks()) { |
There was a problem hiding this comment.
I'm confused, why is this ~=?
| return false; | ||
| } | ||
|
|
||
| std::string filebase_for_block(uint32_t block_num) { |
There was a problem hiding this comment.
I did a search for this, is it used?
| return ""; | ||
| } | ||
|
|
||
| bool set_active_item(uint32_t block_num, mapmode mode = mapmode::readonly) { |
There was a problem hiding this comment.
What do you think about this instead being a method like:
optional<uint64_t> get_position(uint32_t block_num, mapmode mode = ...)
since the active item is always set before retrieving the block position?
| } | ||
| else { | ||
| bfs::remove(old_name); | ||
| wlog("${new_name} already exists, just remove ${old_name}", ("old_name", old_name.string())("new_name", new_name.string())); |
There was a problem hiding this comment.
remove -> removing
| } | ||
|
|
||
| static void rename_bundle(bfs::path orig_path, bfs::path new_path) { | ||
| rename_if_not_exists(orig_path.replace_extension(".log"), new_path.replace_extension(".log")); |
There was a problem hiding this comment.
does it matter to the code after calling this, if the rename fails or succeeds?
There was a problem hiding this comment.
It shouldn't matter. The design was to cover the case when the archive or retaining dir is a network shared directory like (NFS mount), where multiple nodeos can write to it. If the file already exists because another nodeos instance has written to it. We should be able to ignore it.
| rename_bundle(dir / name, new_path); | ||
|
|
||
| if (this->collection.size() >= max_retained_files) { | ||
| auto items_to_erase = |
| file.close(); | ||
| file.open(path.generic_string()); | ||
| EOS_ASSERT(file.size() % sizeof(uint64_t) == 0, Exception, | ||
| "The size of ${file} is not the multiple of sizeof(uint64_t)", ("file", path.generic_string())); |
There was a problem hiding this comment.
"not the multiple" => "not a multiple"
| if (!result && chain_state_log) | ||
| result = chain_state_log->get_block_id(block_num); | ||
|
|
||
| if (!result) { |
There was a problem hiding this comment.
Not a must, but would prefer this being:
if (result) {
return result
}
try ...
Is this being considered ? Maybe make the files under "state-history-archive-dir" be readable by nodeos under read-only mode? Since it would make a lot of sense to keep reading those old files but from a slower storage backend |
Change Description
This PR supports the trace and chain state log sppliting in state history plugin and fixes the problem for context free data pruning for splitted block logs.
Change Type
Select ONE
Consensus Changes
API Changes
Documentation Additions
New option for chain plugin is added
Some new options for state history plugin are added: