Context

For timeseries collections, there are two representations for indexes:

• Logical, user visible representation
• Physical, internal bucketed indexes, accessible with {rawData: true}.

For example:

> db.createCollection("myts", {timeseries:{timeField:'t'}})
> db.myts.createIndex({t: 1})
> db.runCommand({listIndexes: "myts"}).cursor.firstBatch.map(i => i.key)
[ { t: 1 } ]
> db.runCommand({listIndexes: "myts", rawData: true}).cursor.firstBatch.map(i => i.key)
[ {'control.min.t': 1, 'control.max.t': 1 } ]

A logical representation always has a physical representation, but the opposite is not always true. For example consider the physical index {control.closed: 1, control.min.t: 1}:

> db.myts.createIndex({"control.closed": 1, "control.min.t": 1}, {rawData: true})
> db.runCommand({listIndexes: "myts"}).cursor.firstBatch.map(i => i.key)
[ { t: 1 } ]
> db.runCommand({listIndexes: "myts", rawData: true}).cursor.firstBatch.map(i => i.key)
[
   { 'control.min.t': 1, 'control.max.t': 1 }
   { 'control.closed': 1, 'control.min.t': 1 }
]

As we can see, logical listIndexes without rawData:true just omits the index.

Problem Statement/Rationale

Currently mongodump/mongorestore lists and recreate the indexes in logical format. So in the case above, the {control.closed: 1, control.min.t: 1} is absent from the dump and is lost after a restore:

> db.runCommand({listIndexes: "myts", rawData: true}).cursor.firstBatch.map(i => i.key)
[
  { 'control.min.t': 1, 'control.max.t': 1 }
  { 'control.closed': 1, 'control.min.t': 1 }
]
> mongodump
> db.dropDatabase()
> mongorestore
> db.runCommand({listIndexes: "myts", rawData: true}).cursor.firstBatch.map(i => i.key)
[
  { 'control.min.t': 1, 'control.max.t': 1 }
] 

Impact

Generally users should never directly create indexes using rawData:true like the above.
However, they have historically been used internally as a workaround for performance issues, so they may exist in customer clusters.

Proposed fix

mongodump and mongorestore should call listIndexes/createIndexes with rawData: true so that timeseries indexes that do not have a logical representation are correctly preserved by a dump+restore.

Affected versions

• The examples above use the v9.0+ viewless timeseries syntax but this problem also exists in older versions with legacy timeseries, This is a day-0 issue since timeseries was released in MongoDB 5.0.