This repository describes the traces from Twitter's in-memory caching (Twemcache/Pelikan) clusters. The current traces were collected from 54 clusters in Mar 2020. The traces are one-week-long. More details are described in the following paper and blog.
The traces are compressed with zstd, to decompress run zstd -d /path/file.
The decompressed traces are plain text structured as comma-separated columns. Each row represents one request in the following format.
timestamp: the time when the cache receives the request, in secanonymized key: the original key with anonymizationkey size: the size of key in bytesvalue size: the size of value in bytesclient id: the anonymized clients (frontend service) who sends the requestoperation: one of get/gets/set/add/replace/cas/append/prepend/delete/incr/decrTTL: the time-to-live (TTL) of the object set by the client, it is 0 when the request is not a write request.
Note that during key anonymization, we preserve the namespaces, for example, if the anonymized key is nz:u:eeW511W3dcH3de3d15ec, the first two fields nz and u are namespaces, note that the namespaces are not necessarily delimited by :, different workloads use different delimiters with different number of namespaces.
A sample of the traces are attached under samples.
The full traces are large (2.8 TB in compressed form, 14 TB uncompressed), and can be downloaded from the following places.
https://ftp.pdl.cmu.edu/pub/datasets/twemcacheWorkload/open_source
http://iotta.snia.org/tracetypes/17
see storj for how to access (Good for worldwide access, especially Asia and Europe, but not available after Dec 2020)
https://pan.baidu.com/s/1Jm2nAW-UhsjXU6JYoA07LA access code: wcws (Good for Asia access, but UI only has Chinese)
These traces are splitted into smaller files of 1000000000 lines (smaller for SNIA) each and compressed with zstd, so a file with name clusterN.0.zst means this file contains the first 1000000000 requests of cluster N.
Feel free to contact us if you have problem downloading the traces.
For different evaluation purposes, we recommend the following clusters/workloads
-
miss ratio related (admission, eviction): cluster52, cluster17 (low miss ratio), cluster18 (low miss ratio), cluster24, cluster44, cluster45, cluster29.
-
write-heavy workloads: cluster12, cluster15, cluster31, cluster37.
-
TTL-related: mix of small and large TTLs: cluster 52, cluster22, cluster25, cluster11; small TTLs only: cluster18, cluster19, cluster6, cluster7.
others?: feel free to contact us if you are looking for a trace for specific purpose.
We release a computed statistics of each cluster workload under stat/, the latest is here.
This table includes the following fields, each field is the mean value of the metric either from production or from the traces.
The fields include production miss ratio,
workload category (1: storage, 2: computation, 3: transient item), key size, value size, request rate, mean object frequency, one-hit-wonder ratio (%), compulsory miss ratio (%), common TTLs, working set size, operations, Zipf alpha.
- Please join our discussion channel for questions and updates.
- We provide a trace bibliography of papers that have used and/or analyzed the traces, and encourage anybody who publishes one to add it to the bibliography by creating an issue or pull request on GitHub.
We thank Carnegie Mellon University PDL, SNIA and Storj for hosting the traces.
