There is a single table per Perf instance. Each table that contains Traces, OrderedParamSets, and the Source locations.
cbt createtable skia families=V:maxversions=1,S:maxversions=1,D:maxversions=1,I:maxversions=1
Read all the OPS hashes from the android table in the perf-bt instance.
cbt --instance perf-bt read android prefix=@ columns=D:H
Traces are stored by breaking them across Tiles. The size of each Tile will vary depending on considerations like how sparse the data is, for example they may vary from 50 to 8000 commits per Tile.
Tile 1 Tile 2 +---------------------+ +---------------------+ | | | | | | | | | ===== trace 1 ===== | | ===== trace 1 ===== | | | | | | ===== trace 2 ===== | | ===== trace 2 ===== | | | | | | ===== trace 3 ===== | | ===== trace 3 ===== | | | | | | ... | | ... | | | | | | | | | +---------------------+ +---------------------+
Caveats our design has to take into account:
Traces may come and go.
The set of params, and the number of them, can vary across traces.
The Params used to specify tests are set outside of Perf and can change at any time.
Data may be sparse, i.e. a sample point may only arrive every 100 commits.
Key and values for params can be very long. E.g.
Traces are identified by their params, i.e. a trace key looks like:
The number of keys and values may change across tests. See infra/go/query for more information on structured keys.
Since each Tile contains traces for a limited time range, i.e. for only a finite set of commits, then we can track the full set of all keys and params seen across all traces in that tile. That data is stored in an OrderedParamSet. See infra/go/paramtools for more info on OrderedParamSets.
Since we have an OrderedParamSet (OPS) for each Tile we can use that to compress trace ids, storing just the offsets in the OPS. I.e. these keys:
Could be stored as:
Note that since each OPS for each Tile is different the encoding of the same key may change from Tile to Tile.
To make querying faster we store indices for each Tile. They are reverse lookup tables that allow finding every trace id that matches a given param (key=value) pair.
tile1:config:8888 => ,0=1,2=0,3=7, ,0=1,2=1,3=18,5=1,
To query for config=8888 we just load the indices for tile1:config:8888 and then load the traces for all the trace ids we find there. More complex queries require loading more indices then taking intersections and unions of the sets of trace ids they contain.
Finally when we write data we also store the Google Cloud Storage filename where the data came from for each point.
To store all of this in a single BigTable table we have different kinds of rows we store and they use different column families.
Values used in row names:
TileKey = (2^32-1) - (tile number) - With 256 values per tile this lets us store 1 trillion values per trace. - Formatted as %07d - Note that this reverses the order of the tiles, i.e. new tiles have smaller numbers, so that we can do a simple query to find the newest tile. TraceKey = OrderedParamSet.EncodeParamsAsString() - A structured key using just the offsets, e.g. ",0=1,2=102,3=1," Shard = A number, calculated from the TraceKey, that places the trace in one of the shards. The total number of shards is set per table.
V - Column family stores float32 values - Columns: 0, 1, 2, ..., TILE_SIZE-1 S - Column family stores md5 sum of the source location, written as byte. Look up the actual value of the source under the H column family. - Columns: 0, 1, 2, ..., TILE_SIZE-1
D - Column family stores OrderedParamSets. - Columns: R - Revision (hash of stored OPS to avoid the lost update problem). OPS - The serialized Ordered ParamSet
row name = ‘j’ + TileKey:ParamKey:ParamValue:traceID (A previous indexing scheme used ‘i’ as a prefix, so we avoid using that prefix.)
I - Column family stores the empty string. - Columns: E - An empty byte slice. (You can't create a row w/o writing to at least one column, so we store the smallest amount of data possible since all the data is in the rowname.)
H - Column family stores md5 hash of source file name written as hex string. - Columns: S - Source (The full name of the source file, gs://....)
We keep indices for each Tile for each key=value pair that appears in any trace ID. These indices can be quite large and can't be stored in memory, so querying has to be done in a streaming manner. Each box
+---------+ |key=value| +---------+
in the diagram below represents a query against BigTable against the indices for that key-value pair, and the arrow out is a channel that provides the trace IDs for all traces that match that query, provided in ascending order with no duplicates. Any query in Perf can be boiled down to just a union of queries across matching keys, and an intersection of queries across different keys.
^ | +--+--+ | ∩ | +-----+ ^ ^ | | +-----------+ +--------+ | | +--+--+ | | ∪ | | +-----+ | ^ ^ | | | | +----------++ ++---------+ +----+---+ |config=8888| |config=565| |arch=x86| +-----------+ +----------+ +--------+
The ∪ and ∩ nodes are running Go routines that take in N incoming channels of strings in order and produce either the union or the intersection respectively of the incoming channel. This allows the amount of memory used to be kept to a minimum while providing a stream of trace ids that match a query, which can incrementally be bundled into requests back to BT to retrieve the actual trace values.
Note that while we use normal keys and values in the example above, the implementation will use encoded keys and values for space savings.
There is a command line tool to inspect aspects of the BigTable store.
The --big_table_config flag takes the name of a known config for Perf. See the file go/config/config.go for the full list of configs.
You can find the index of the most recent tile:
perf-tool tiles last
Or force the index for a tile to be re-written:
perf-tool --logtostderr indices write --big_table_config=nano
Or try queries:
perf-tool traces list-by-index --query=name=A_large_blank_world_map_with_oceans_marked_in_blue.svg\&sub_result=min_ms